{"arms":{"direct":{"math":{"accuracy":0.706989247311828,"category":{"Algebra":{"accuracy":0.8260869565217391,"n":92},"Counting & Probability":{"accuracy":0.6428571428571429,"n":28},"Geometry":{"accuracy":0.5806451612903226,"n":31},"Intermediate Algebra":{"accuracy":0.5138888888888888,"n":72},"Number Theory":{"accuracy":0.782608695652174,"n":46},"Prealgebra":{"accuracy":0.8524590163934426,"n":61},"Precalculus":{"accuracy":0.6190476190476191,"n":42}},"level":{"1":{"accuracy":0.875,"n":32},"2":{"accuracy":0.8507462686567164,"n":67},"3":{"accuracy":0.8846153846153846,"n":78},"4":{"accuracy":0.6421052631578947,"n":95},"5":{"accuracy":0.48,"n":100}},"n":372,"truncation_rate":0.024193548387096774},"retention":{"accuracy":0.42578125,"category":{"biology":{"accuracy":0.8235294117647058,"n":34},"business":{"accuracy":0.21052631578947367,"n":38},"chemistry":{"accuracy":0.2777777777777778,"n":54},"computer science":{"accuracy":0.45,"n":20},"economics":{"accuracy":0.6097560975609756,"n":41},"engineering":{"accuracy":0.23404255319148937,"n":47},"health":{"accuracy":0.5897435897435898,"n":39},"history":{"accuracy":0.5,"n":18},"law":{"accuracy":0.41509433962264153,"n":53},"other":{"accuracy":0.45454545454545453,"n":44},"philosophy":{"accuracy":0.4166666666666667,"n":24},"physics":{"accuracy":0.25806451612903225,"n":62},"psychology":{"accuracy":0.5789473684210527,"n":38}},"level":{"":{"accuracy":0.42578125,"n":512}},"n":512,"truncation_rate":0.0}},"naive":{"math":{"accuracy":0.8064516129032258,"category":{"Algebra":{"accuracy":0.9782608695652174,"n":92},"Counting & Probability":{"accuracy":0.7857142857142857,"n":28},"Geometry":{"accuracy":0.5161290322580645,"n":31},"Intermediate Algebra":{"accuracy":0.5972222222222222,"n":72},"Number Theory":{"accuracy":0.9130434782608695,"n":46},"Prealgebra":{"accuracy":0.8852459016393442,"n":61},"Precalculus":{"accuracy":0.7857142857142857,"n":42}},"level":{"1":{"accuracy":0.96875,"n":32},"2":{"accuracy":0.9552238805970149,"n":67},"3":{"accuracy":0.9358974358974359,"n":78},"4":{"accuracy":0.7578947368421053,"n":95},"5":{"accuracy":0.6,"n":100}},"n":372,"truncation_rate":0.013440860215053764},"retention":{"accuracy":0.494140625,"category":{"biology":{"accuracy":0.8529411764705882,"n":34},"business":{"accuracy":0.2631578947368421,"n":38},"chemistry":{"accuracy":0.3888888888888889,"n":54},"computer science":{"accuracy":0.55,"n":20},"economics":{"accuracy":0.7560975609756098,"n":41},"engineering":{"accuracy":0.40425531914893614,"n":47},"health":{"accuracy":0.5897435897435898,"n":39},"history":{"accuracy":0.5555555555555556,"n":18},"law":{"accuracy":0.4339622641509434,"n":53},"other":{"accuracy":0.4772727272727273,"n":44},"philosophy":{"accuracy":0.625,"n":24},"physics":{"accuracy":0.27419354838709675,"n":62},"psychology":{"accuracy":0.6052631578947368,"n":38}},"level":{"":{"accuracy":0.494140625,"n":512}},"n":512,"truncation_rate":0.0}},"selected":{"math":{"accuracy":0.8225806451612904,"category":{"Algebra":{"accuracy":0.9456521739130435,"n":92},"Counting & Probability":{"accuracy":0.75,"n":28},"Geometry":{"accuracy":0.6129032258064516,"n":31},"Intermediate Algebra":{"accuracy":0.6111111111111112,"n":72},"Number Theory":{"accuracy":0.9782608695652174,"n":46},"Prealgebra":{"accuracy":0.9344262295081968,"n":61},"Precalculus":{"accuracy":0.7857142857142857,"n":42}},"level":{"1":{"accuracy":0.96875,"n":32},"2":{"accuracy":0.9104477611940298,"n":67},"3":{"accuracy":0.9230769230769231,"n":78},"4":{"accuracy":0.8421052631578947,"n":95},"5":{"accuracy":0.62,"n":100}},"n":372,"truncation_rate":0.010752688172043012},"retention":{"accuracy":0.501953125,"category":{"biology":{"accuracy":0.8529411764705882,"n":34},"business":{"accuracy":0.2631578947368421,"n":38},"chemistry":{"accuracy":0.3888888888888889,"n":54},"computer science":{"accuracy":0.55,"n":20},"economics":{"accuracy":0.7560975609756098,"n":41},"engineering":{"accuracy":0.425531914893617,"n":47},"health":{"accuracy":0.5897435897435898,"n":39},"history":{"accuracy":0.5555555555555556,"n":18},"law":{"accuracy":0.4339622641509434,"n":53},"other":{"accuracy":0.5,"n":44},"philosophy":{"accuracy":0.5833333333333334,"n":24},"physics":{"accuracy":0.3225806451612903,"n":62},"psychology":{"accuracy":0.6052631578947368,"n":38}},"level":{"":{"accuracy":0.501953125,"n":512}},"n":512,"truncation_rate":0.0}},"untouched":{"math":{"accuracy":0.8360215053763441,"category":{"Algebra":{"accuracy":0.967391304347826,"n":92},"Counting & Probability":{"accuracy":0.7857142857142857,"n":28},"Geometry":{"accuracy":0.6774193548387096,"n":31},"Intermediate Algebra":{"accuracy":0.6666666666666666,"n":72},"Number Theory":{"accuracy":0.9347826086956522,"n":46},"Prealgebra":{"accuracy":0.8852459016393442,"n":61},"Precalculus":{"accuracy":0.8095238095238095,"n":42}},"level":{"1":{"accuracy":1.0,"n":32},"2":{"accuracy":0.9402985074626866,"n":67},"3":{"accuracy":0.9487179487179487,"n":78},"4":{"accuracy":0.8,"n":95},"5":{"accuracy":0.66,"n":100}},"n":372,"truncation_rate":0.005376344086021506},"retention":{"accuracy":0.498046875,"category":{"biology":{"accuracy":0.8529411764705882,"n":34},"business":{"accuracy":0.2894736842105263,"n":38},"chemistry":{"accuracy":0.35185185185185186,"n":54},"computer science":{"accuracy":0.55,"n":20},"economics":{"accuracy":0.7073170731707317,"n":41},"engineering":{"accuracy":0.44680851063829785,"n":47},"health":{"accuracy":0.5897435897435898,"n":39},"history":{"accuracy":0.5555555555555556,"n":18},"law":{"accuracy":0.41509433962264153,"n":53},"other":{"accuracy":0.5,"n":44},"philosophy":{"accuracy":0.5833333333333334,"n":24},"physics":{"accuracy":0.3548387096774194,"n":62},"psychology":{"accuracy":0.5789473684210527,"n":38}},"level":{"":{"accuracy":0.498046875,"n":512}},"n":512,"truncation_rate":0.0}}},"score":{"best_control":0.8360215053763441,"retention_pass":{"direct":false,"naive":true,"selected":true,"untouched":true},"reward":0.0,"selected_minus_direct":0.11559139784946237,"selected_minus_naive":0.016129032258064613,"selected_minus_untouched":-0.013440860215053752},"paired_contrasts":{"direct":{"difference":0.11559139784946236,"losses":9,"n":372,"paired_bootstrap_95pct":[0.07795698924731183,0.15591397849462366],"wins":52},"naive":{"difference":0.016129032258064516,"losses":13,"n":372,"paired_bootstrap_95pct":[-0.013440860215053764,0.0456989247311828],"wins":19},"untouched":{"difference":-0.013440860215053764,"losses":24,"n":372,"paired_bootstrap_95pct":[-0.04838709677419355,0.021505376344086023],"wins":19}},"budget":{"compliant":true,"proxy_experiment_gpu_seconds":5216.889497671975,"proxy_positions":4186112},"proxy_trials":[{"diagnostics":{"eligibility":{"reference":510,"teacher":0},"presentations":256,"teacher_presentations":0,"unique_prompts":256},"id":"p0-reference-mid","metrics":{"math":{"accuracy":0.265625,"category":{"Algebra":{"accuracy":0.4117647058823529,"n":17},"Counting & Probability":{"accuracy":0.2857142857142857,"n":7},"Geometry":{"accuracy":0.5,"n":6},"Intermediate Algebra":{"accuracy":0.14285714285714285,"n":14},"Number Theory":{"accuracy":0.0,"n":6},"Prealgebra":{"accuracy":0.3333333333333333,"n":9},"Precalculus":{"accuracy":0.0,"n":5}},"level":{"1":{"accuracy":1.0,"n":5},"2":{"accuracy":0.45454545454545453,"n":11},"3":{"accuracy":0.3333333333333333,"n":15},"4":{"accuracy":0.125,"n":16},"5":{"accuracy":0.0,"n":17}},"n":64,"truncation_rate":0.125},"retention":{"accuracy":0.2265625,"category":{"biology":{"accuracy":0.3333333333333333,"n":9},"business":{"accuracy":0.0,"n":10},"chemistry":{"accuracy":0.07142857142857142,"n":14},"computer science":{"accuracy":0.4,"n":5},"economics":{"accuracy":0.4,"n":10},"engineering":{"accuracy":0.16666666666666666,"n":12},"health":{"accuracy":0.3,"n":10},"history":{"accuracy":0.0,"n":4},"law":{"accuracy":0.15384615384615385,"n":13},"other":{"accuracy":0.2727272727272727,"n":11},"philosophy":{"accuracy":0.16666666666666666,"n":6},"physics":{"accuracy":0.13333333333333333,"n":15},"psychology":{"accuracy":0.6666666666666666,"n":9}},"level":{"":{"accuracy":0.2265625,"n":128}},"n":128,"truncation_rate":0.0}},"seconds":445.1589773320593,"training":{"adapter_sha256":"20b4cce7912903ef757990a6ff48f105baee00bf8c4a73bcc709a20d6a2ce760","peak_allocated_bytes":11305907200,"position_limit":524288,"positions":524288,"recipe":{"accumulation":4,"alpha":32,"beta":0.02,"group_size":4,"learning_rate":0.0001,"method":"reference_sft","micro_batch":2,"rank":16,"rollout_tokens":1024,"seed":7107,"teacher_fraction":0.0,"temperature":0.8},"role":"proxy","teacher_cache_hash":null,"train_seconds":35.04000649298541,"training_data_hash":"1118fbf8abc66267f6e0d05201f62f9d4a9b13ee6a8a9626aa4d04ec03aa1284","updates":32}},{"diagnostics":{"eligibility":{"reference":510,"teacher":0},"presentations":256,"teacher_presentations":0,"unique_prompts":256},"id":"p1-reference-high","metrics":{"math":{"accuracy":0.171875,"category":{"Algebra":{"accuracy":0.17647058823529413,"n":17},"Counting & Probability":{"accuracy":0.14285714285714285,"n":7},"Geometry":{"accuracy":0.16666666666666666,"n":6},"Intermediate Algebra":{"accuracy":0.21428571428571427,"n":14},"Number Theory":{"accuracy":0.0,"n":6},"Prealgebra":{"accuracy":0.3333333333333333,"n":9},"Precalculus":{"accuracy":0.0,"n":5}},"level":{"1":{"accuracy":0.6,"n":5},"2":{"accuracy":0.36363636363636365,"n":11},"3":{"accuracy":0.13333333333333333,"n":15},"4":{"accuracy":0.0625,"n":16},"5":{"accuracy":0.058823529411764705,"n":17}},"n":64,"truncation_rate":0.078125},"retention":{"accuracy":0.234375,"category":{"biology":{"accuracy":0.2222222222222222,"n":9},"business":{"accuracy":0.0,"n":10},"chemistry":{"accuracy":0.14285714285714285,"n":14},"computer science":{"accuracy":0.2,"n":5},"economics":{"accuracy":0.4,"n":10},"engineering":{"accuracy":0.25,"n":12},"health":{"accuracy":0.3,"n":10},"history":{"accuracy":0.0,"n":4},"law":{"accuracy":0.23076923076923078,"n":13},"other":{"accuracy":0.2727272727272727,"n":11},"philosophy":{"accuracy":0.16666666666666666,"n":6},"physics":{"accuracy":0.13333333333333333,"n":15},"psychology":{"accuracy":0.6666666666666666,"n":9}},"level":{"":{"accuracy":0.234375,"n":128}},"n":128,"truncation_rate":0.0}},"seconds":441.52898574783467,"training":{"adapter_sha256":"b5293290453cdf4854e9a155cedaec44b503077cbf007cf2a38f604b61e7ea27","peak_allocated_bytes":11305907200,"position_limit":524288,"positions":524288,"recipe":{"accumulation":4,"alpha":32,"beta":0.02,"group_size":4,"learning_rate":0.0003,"method":"reference_sft","micro_batch":2,"rank":16,"rollout_tokens":1024,"seed":7107,"teacher_fraction":0.0,"temperature":0.8},"role":"proxy","teacher_cache_hash":null,"train_seconds":37.87431682390161,"training_data_hash":"1118fbf8abc66267f6e0d05201f62f9d4a9b13ee6a8a9626aa4d04ec03aa1284","updates":32}},{"diagnostics":{"eligibility":{"reference":510,"teacher":0},"presentations":256,"teacher_presentations":0,"unique_prompts":256},"id":"p2-reference-low","metrics":{"math":{"accuracy":0.21875,"category":{"Algebra":{"accuracy":0.29411764705882354,"n":17},"Counting & Probability":{"accuracy":0.2857142857142857,"n":7},"Geometry":{"accuracy":0.16666666666666666,"n":6},"Intermediate Algebra":{"accuracy":0.14285714285714285,"n":14},"Number Theory":{"accuracy":0.0,"n":6},"Prealgebra":{"accuracy":0.4444444444444444,"n":9},"Precalculus":{"accuracy":0.0,"n":5}},"level":{"1":{"accuracy":0.6,"n":5},"2":{"accuracy":0.2727272727272727,"n":11},"3":{"accuracy":0.4,"n":15},"4":{"accuracy":0.125,"n":16},"5":{"accuracy":0.0,"n":17}},"n":64,"truncation_rate":0.09375},"retention":{"accuracy":0.2578125,"category":{"biology":{"accuracy":0.3333333333333333,"n":9},"business":{"accuracy":0.0,"n":10},"chemistry":{"accuracy":0.07142857142857142,"n":14},"computer science":{"accuracy":0.6,"n":5},"economics":{"accuracy":0.5,"n":10},"engineering":{"accuracy":0.25,"n":12},"health":{"accuracy":0.4,"n":10},"history":{"accuracy":0.0,"n":4},"law":{"accuracy":0.15384615384615385,"n":13},"other":{"accuracy":0.2727272727272727,"n":11},"philosophy":{"accuracy":0.16666666666666666,"n":6},"physics":{"accuracy":0.13333333333333333,"n":15},"psychology":{"accuracy":0.6666666666666666,"n":9}},"level":{"":{"accuracy":0.2578125,"n":128}},"n":128,"truncation_rate":0.0}},"seconds":379.79951210389845,"training":{"adapter_sha256":"1e7ff45689cbb4a410dcda9b7457564a800f69f8e485fe9d514be409a6c55159","peak_allocated_bytes":11305907200,"position_limit":524288,"positions":524288,"recipe":{"accumulation":4,"alpha":32,"beta":0.02,"group_size":4,"learning_rate":3e-05,"method":"reference_sft","micro_batch":2,"rank":16,"rollout_tokens":1024,"seed":7107,"teacher_fraction":0.0,"temperature":0.8},"role":"proxy","teacher_cache_hash":null,"train_seconds":34.86071789520793,"training_data_hash":"1118fbf8abc66267f6e0d05201f62f9d4a9b13ee6a8a9626aa4d04ec03aa1284","updates":32}},{"diagnostics":{"eligibility":{"reference":510,"teacher":275},"presentations":256,"teacher_presentations":256,"unique_prompts":256},"id":"p3-teacher","metrics":{"math":{"accuracy":0.375,"category":{"Algebra":{"accuracy":0.4117647058823529,"n":17},"Counting & Probability":{"accuracy":0.42857142857142855,"n":7},"Geometry":{"accuracy":0.3333333333333333,"n":6},"Intermediate Algebra":{"accuracy":0.14285714285714285,"n":14},"Number Theory":{"accuracy":0.5,"n":6},"Prealgebra":{"accuracy":0.5555555555555556,"n":9},"Precalculus":{"accuracy":0.4,"n":5}},"level":{"1":{"accuracy":0.8,"n":5},"2":{"accuracy":0.5454545454545454,"n":11},"3":{"accuracy":0.6,"n":15},"4":{"accuracy":0.25,"n":16},"5":{"accuracy":0.058823529411764705,"n":17}},"n":64,"truncation_rate":0.140625},"retention":{"accuracy":0.25,"category":{"biology":{"accuracy":0.3333333333333333,"n":9},"business":{"accuracy":0.0,"n":10},"chemistry":{"accuracy":0.07142857142857142,"n":14},"computer science":{"accuracy":0.6,"n":5},"economics":{"accuracy":0.5,"n":10},"engineering":{"accuracy":0.25,"n":12},"health":{"accuracy":0.3,"n":10},"history":{"accuracy":0.0,"n":4},"law":{"accuracy":0.15384615384615385,"n":13},"other":{"accuracy":0.2727272727272727,"n":11},"philosophy":{"accuracy":0.16666666666666666,"n":6},"physics":{"accuracy":0.13333333333333333,"n":15},"psychology":{"accuracy":0.6666666666666666,"n":9}},"level":{"":{"accuracy":0.25,"n":128}},"n":128,"truncation_rate":0.0}},"seconds":454.6644635749981,"training":{"adapter_sha256":"78759c075f7bcfe06a8185278fae8f996bcc4edda89f93d07181b190ccb9bb2f","peak_allocated_bytes":11305907200,"position_limit":524288,"positions":524288,"recipe":{"accumulation":4,"alpha":32,"beta":0.02,"group_size":4,"learning_rate":0.0001,"method":"teacher_sft","micro_batch":2,"rank":16,"rollout_tokens":1024,"seed":7107,"teacher_fraction":1.0,"temperature":0.8},"role":"proxy","teacher_cache_hash":"6629bea9b6dd9c749d7613305a51f461a8c861be74f96ad4dc072e027759e1ba","train_seconds":38.58267200598493,"training_data_hash":"1118fbf8abc66267f6e0d05201f62f9d4a9b13ee6a8a9626aa4d04ec03aa1284","updates":32}},{"diagnostics":{"eligibility":{"reference":510,"teacher":275},"presentations":256,"teacher_presentations":128,"unique_prompts":220},"id":"p4-mixed","metrics":{"math":{"accuracy":0.390625,"category":{"Algebra":{"accuracy":0.47058823529411764,"n":17},"Counting & Probability":{"accuracy":0.5714285714285714,"n":7},"Geometry":{"accuracy":0.0,"n":6},"Intermediate Algebra":{"accuracy":0.21428571428571427,"n":14},"Number Theory":{"accuracy":0.3333333333333333,"n":6},"Prealgebra":{"accuracy":0.6666666666666666,"n":9},"Precalculus":{"accuracy":0.4,"n":5}},"level":{"1":{"accuracy":1.0,"n":5},"2":{"accuracy":0.45454545454545453,"n":11},"3":{"accuracy":0.6,"n":15},"4":{"accuracy":0.3125,"n":16},"5":{"accuracy":0.058823529411764705,"n":17}},"n":64,"truncation_rate":0.125},"retention":{"accuracy":0.234375,"category":{"biology":{"accuracy":0.2222222222222222,"n":9},"business":{"accuracy":0.0,"n":10},"chemistry":{"accuracy":0.14285714285714285,"n":14},"computer science":{"accuracy":0.4,"n":5},"economics":{"accuracy":0.4,"n":10},"engineering":{"accuracy":0.16666666666666666,"n":12},"health":{"accuracy":0.4,"n":10},"history":{"accuracy":0.0,"n":4},"law":{"accuracy":0.15384615384615385,"n":13},"other":{"accuracy":0.2727272727272727,"n":11},"philosophy":{"accuracy":0.16666666666666666,"n":6},"physics":{"accuracy":0.13333333333333333,"n":15},"psychology":{"accuracy":0.6666666666666666,"n":9}},"level":{"":{"accuracy":0.234375,"n":128}},"n":128,"truncation_rate":0.0}},"seconds":459.6238218541257,"training":{"adapter_sha256":"5b4db0af97bd2a754fb52bd2363f0702ec444d4dd30a8ec55f23a0e7ad2bfd0d","peak_allocated_bytes":11305907200,"position_limit":524288,"positions":524288,"recipe":{"accumulation":4,"alpha":32,"beta":0.02,"group_size":4,"learning_rate":0.0001,"method":"mixed_sft","micro_batch":2,"rank":16,"rollout_tokens":1024,"seed":7107,"teacher_fraction":0.5,"temperature":0.8},"role":"proxy","teacher_cache_hash":"6629bea9b6dd9c749d7613305a51f461a8c861be74f96ad4dc072e027759e1ba","train_seconds":38.62453570798971,"training_data_hash":"1118fbf8abc66267f6e0d05201f62f9d4a9b13ee6a8a9626aa4d04ec03aa1284","updates":32}},{"diagnostics":{"generated_tokens":46138,"groups":42,"responses":84,"truncated_responses":2,"unique_prompts":42},"id":"p5-opd","metrics":{"math":{"accuracy":0.359375,"category":{"Algebra":{"accuracy":0.4117647058823529,"n":17},"Counting & Probability":{"accuracy":0.2857142857142857,"n":7},"Geometry":{"accuracy":0.16666666666666666,"n":6},"Intermediate Algebra":{"accuracy":0.21428571428571427,"n":14},"Number Theory":{"accuracy":0.5,"n":6},"Prealgebra":{"accuracy":0.6666666666666666,"n":9},"Precalculus":{"accuracy":0.2,"n":5}},"level":{"1":{"accuracy":1.0,"n":5},"2":{"accuracy":0.45454545454545453,"n":11},"3":{"accuracy":0.6666666666666666,"n":15},"4":{"accuracy":0.125,"n":16},"5":{"accuracy":0.058823529411764705,"n":17}},"n":64,"truncation_rate":0.140625},"retention":{"accuracy":0.28125,"category":{"biology":{"accuracy":0.3333333333333333,"n":9},"business":{"accuracy":0.1,"n":10},"chemistry":{"accuracy":0.21428571428571427,"n":14},"computer science":{"accuracy":0.6,"n":5},"economics":{"accuracy":0.5,"n":10},"engineering":{"accuracy":0.25,"n":12},"health":{"accuracy":0.4,"n":10},"history":{"accuracy":0.0,"n":4},"law":{"accuracy":0.15384615384615385,"n":13},"other":{"accuracy":0.2727272727272727,"n":11},"philosophy":{"accuracy":0.16666666666666666,"n":6},"physics":{"accuracy":0.13333333333333333,"n":15},"psychology":{"accuracy":0.6666666666666666,"n":9}},"level":{"":{"accuracy":0.28125,"n":128}},"n":128,"truncation_rate":0.0}},"seconds":1416.0745393850375,"training":{"adapter_sha256":"1369a5be794086cc62714d17c689e1b6293b5cfdb30426c3eeb4a1844d5906aa","peak_allocated_bytes":23784590336,"position_limit":524288,"positions":516096,"recipe":{"accumulation":4,"alpha":32,"beta":0.02,"group_size":2,"learning_rate":3e-05,"method":"opd","micro_batch":2,"rank":16,"rollout_tokens":1024,"seed":7107,"teacher_fraction":0.0,"temperature":0.8},"role":"proxy","teacher_cache_hash":null,"train_seconds":979.4048202189151,"training_data_hash":"1118fbf8abc66267f6e0d05201f62f9d4a9b13ee6a8a9626aa4d04ec03aa1284","updates":42}},{"diagnostics":{"generated_tokens":37646,"groups":16,"mean_reward":0.234375,"mixed_reward_groups":6,"responses":64,"truncated_responses":2,"unique_prompts":16},"id":"p6-grpo-low","metrics":{"math":{"accuracy":0.359375,"category":{"Algebra":{"accuracy":0.29411764705882354,"n":17},"Counting & Probability":{"accuracy":0.42857142857142855,"n":7},"Geometry":{"accuracy":0.0,"n":6},"Intermediate Algebra":{"accuracy":0.35714285714285715,"n":14},"Number Theory":{"accuracy":0.8333333333333334,"n":6},"Prealgebra":{"accuracy":0.5555555555555556,"n":9},"Precalculus":{"accuracy":0.0,"n":5}},"level":{"1":{"accuracy":0.8,"n":5},"2":{"accuracy":0.45454545454545453,"n":11},"3":{"accuracy":0.4666666666666667,"n":15},"4":{"accuracy":0.25,"n":16},"5":{"accuracy":0.17647058823529413,"n":17}},"n":64,"truncation_rate":0.109375},"retention":{"accuracy":0.296875,"category":{"biology":{"accuracy":0.3333333333333333,"n":9},"business":{"accuracy":0.1,"n":10},"chemistry":{"accuracy":0.21428571428571427,"n":14},"computer science":{"accuracy":0.6,"n":5},"economics":{"accuracy":0.5,"n":10},"engineering":{"accuracy":0.3333333333333333,"n":12},"health":{"accuracy":0.4,"n":10},"history":{"accuracy":0.25,"n":4},"law":{"accuracy":0.15384615384615385,"n":13},"other":{"accuracy":0.2727272727272727,"n":11},"philosophy":{"accuracy":0.16666666666666666,"n":6},"physics":{"accuracy":0.13333333333333333,"n":15},"psychology":{"accuracy":0.6666666666666666,"n":9}},"level":{"":{"accuracy":0.296875,"n":128}},"n":128,"truncation_rate":0.0}},"seconds":863.5845693750307,"training":{"adapter_sha256":"d642e834883e9d6eda29b494a9ae10b3d1e27dc734b37250a21d1346aec74527","peak_allocated_bytes":13638536704,"position_limit":524288,"positions":524288,"recipe":{"accumulation":4,"alpha":32,"beta":0.02,"group_size":4,"learning_rate":3e-05,"method":"grpo","micro_batch":2,"rank":16,"rollout_tokens":1024,"seed":7107,"teacher_fraction":0.0,"temperature":0.8},"role":"proxy","teacher_cache_hash":null,"train_seconds":419.01613774499856,"training_data_hash":"1118fbf8abc66267f6e0d05201f62f9d4a9b13ee6a8a9626aa4d04ec03aa1284","updates":16}},{"diagnostics":{"generated_tokens":33949,"groups":16,"mean_reward":0.203125,"mixed_reward_groups":6,"responses":64,"truncated_responses":3,"unique_prompts":16},"id":"p7-grpo-high","metrics":{"math":{"accuracy":0.359375,"category":{"Algebra":{"accuracy":0.47058823529411764,"n":17},"Counting & Probability":{"accuracy":0.5714285714285714,"n":7},"Geometry":{"accuracy":0.0,"n":6},"Intermediate Algebra":{"accuracy":0.21428571428571427,"n":14},"Number Theory":{"accuracy":0.3333333333333333,"n":6},"Prealgebra":{"accuracy":0.6666666666666666,"n":9},"Precalculus":{"accuracy":0.0,"n":5}},"level":{"1":{"accuracy":1.0,"n":5},"2":{"accuracy":0.45454545454545453,"n":11},"3":{"accuracy":0.4666666666666667,"n":15},"4":{"accuracy":0.1875,"n":16},"5":{"accuracy":0.17647058823529413,"n":17}},"n":64,"truncation_rate":0.125},"retention":{"accuracy":0.2734375,"category":{"biology":{"accuracy":0.3333333333333333,"n":9},"business":{"accuracy":0.0,"n":10},"chemistry":{"accuracy":0.21428571428571427,"n":14},"computer science":{"accuracy":0.6,"n":5},"economics":{"accuracy":0.5,"n":10},"engineering":{"accuracy":0.25,"n":12},"health":{"accuracy":0.4,"n":10},"history":{"accuracy":0.0,"n":4},"law":{"accuracy":0.15384615384615385,"n":13},"other":{"accuracy":0.2727272727272727,"n":11},"philosophy":{"accuracy":0.16666666666666666,"n":6},"physics":{"accuracy":0.13333333333333333,"n":15},"psychology":{"accuracy":0.6666666666666666,"n":9}},"level":{"":{"accuracy":0.2734375,"n":128}},"n":128,"truncation_rate":0.0}},"seconds":756.4546282989904,"training":{"adapter_sha256":"47103ee68b1136eec763025d18701babd46b2a23c8b218f32ea228ceadfbc756","peak_allocated_bytes":13638536704,"position_limit":524288,"positions":524288,"recipe":{"accumulation":4,"alpha":32,"beta":0.02,"group_size":4,"learning_rate":0.0001,"method":"grpo","micro_batch":2,"rank":16,"rollout_tokens":1024,"seed":7107,"teacher_fraction":0.0,"temperature":0.8},"role":"proxy","teacher_cache_hash":null,"train_seconds":348.0023534731008,"training_data_hash":"1118fbf8abc66267f6e0d05201f62f9d4a9b13ee6a8a9626aa4d04ec03aa1284","updates":16}}],"target_receipts":{"direct":{"adapter_sha256":"81a13e25154c276dd2ac42b9e60ccad4566d907862ae0434bb6b8c6bb47a6524","peak_allocated_bytes":40442245120,"position_limit":524288,"positions":524288,"recipe":{"accumulation":4,"alpha":32,"beta":0.02,"group_size":4,"learning_rate":0.0001,"method":"reference_sft","micro_batch":2,"rank":16,"rollout_tokens":1024,"seed":7107,"teacher_fraction":0.0,"temperature":0.8},"role":"target","teacher_cache_hash":null,"train_seconds":107.60353357088752,"training_data_hash":"1118fbf8abc66267f6e0d05201f62f9d4a9b13ee6a8a9626aa4d04ec03aa1284","updates":32},"naive":{"adapter_sha256":"c963b45890ecb83fb8c730bd2d2ef29517aef6ff32e0431ac5a097796c8fcc65","peak_allocated_bytes":52718982144,"position_limit":524288,"positions":516096,"recipe":{"accumulation":4,"alpha":32,"beta":0.02,"group_size":2,"learning_rate":3e-05,"method":"opd","micro_batch":2,"rank":16,"rollout_tokens":1024,"seed":7107,"teacher_fraction":0.0,"temperature":0.8},"role":"target","teacher_cache_hash":null,"train_seconds":1684.7486601730343,"training_data_hash":"1118fbf8abc66267f6e0d05201f62f9d4a9b13ee6a8a9626aa4d04ec03aa1284","updates":42},"selected":{"adapter_sha256":"5f160ae39652cc6abe890f87beef90b9896f736af933101f331eea8d76b9be98","peak_allocated_bytes":52718982144,"position_limit":524288,"positions":516096,"recipe":{"accumulation":4,"alpha":32,"beta":0.02,"group_size":2,"learning_rate":9.819805060619657e-06,"method":"opd","micro_batch":2,"rank":16,"rollout_tokens":1024,"seed":7107,"teacher_fraction":0.0,"temperature":0.8},"role":"target","teacher_cache_hash":null,"train_seconds":1676.2896842889022,"training_data_hash":"1118fbf8abc66267f6e0d05201f62f9d4a9b13ee6a8a9626aa4d04ec03aa1284","updates":42}},"target_diagnostics":{"direct":{"eligibility":{"reference":510,"teacher":0},"presentations":256,"teacher_presentations":0,"unique_prompts":256},"naive":{"generated_tokens":54228,"groups":42,"responses":84,"truncated_responses":11,"unique_prompts":42},"selected":{"generated_tokens":54192,"groups":42,"responses":84,"truncated_responses":15,"unique_prompts":42}},"prechecks":{"metrics":{"proxy":{"math":{"accuracy":0.359375,"category":{"Algebra":{"accuracy":0.4117647058823529,"n":17},"Counting & Probability":{"accuracy":0.42857142857142855,"n":7},"Geometry":{"accuracy":0.0,"n":6},"Intermediate Algebra":{"accuracy":0.2857142857142857,"n":14},"Number Theory":{"accuracy":0.3333333333333333,"n":6},"Prealgebra":{"accuracy":0.6666666666666666,"n":9},"Precalculus":{"accuracy":0.2,"n":5}},"level":{"1":{"accuracy":0.8,"n":5},"2":{"accuracy":0.5454545454545454,"n":11},"3":{"accuracy":0.6,"n":15},"4":{"accuracy":0.1875,"n":16},"5":{"accuracy":0.058823529411764705,"n":17}},"n":64,"truncation_rate":0.109375},"retention":{"accuracy":0.296875,"category":{"biology":{"accuracy":0.3333333333333333,"n":9},"business":{"accuracy":0.1,"n":10},"chemistry":{"accuracy":0.21428571428571427,"n":14},"computer science":{"accuracy":0.6,"n":5},"economics":{"accuracy":0.5,"n":10},"engineering":{"accuracy":0.3333333333333333,"n":12},"health":{"accuracy":0.4,"n":10},"history":{"accuracy":0.0,"n":4},"law":{"accuracy":0.15384615384615385,"n":13},"other":{"accuracy":0.2727272727272727,"n":11},"philosophy":{"accuracy":0.16666666666666666,"n":6},"physics":{"accuracy":0.13333333333333333,"n":15},"psychology":{"accuracy":0.7777777777777778,"n":9}},"level":{"":{"accuracy":0.296875,"n":128}},"n":128,"truncation_rate":0.0}},"target":{"math":{"accuracy":0.828125,"category":{"Algebra":{"accuracy":1.0,"n":17},"Counting & Probability":{"accuracy":0.8571428571428571,"n":7},"Geometry":{"accuracy":0.6666666666666666,"n":6},"Intermediate Algebra":{"accuracy":0.5714285714285714,"n":14},"Number Theory":{"accuracy":1.0,"n":6},"Prealgebra":{"accuracy":0.8888888888888888,"n":9},"Precalculus":{"accuracy":0.8,"n":5}},"level":{"1":{"accuracy":1.0,"n":5},"2":{"accuracy":1.0,"n":11},"3":{"accuracy":1.0,"n":15},"4":{"accuracy":0.875,"n":16},"5":{"accuracy":0.47058823529411764,"n":17}},"n":64,"truncation_rate":0.0},"retention":{"accuracy":0.5078125,"category":{"biology":{"accuracy":0.8888888888888888,"n":9},"business":{"accuracy":0.2,"n":10},"chemistry":{"accuracy":0.35714285714285715,"n":14},"computer science":{"accuracy":0.8,"n":5},"economics":{"accuracy":0.5,"n":10},"engineering":{"accuracy":0.3333333333333333,"n":12},"health":{"accuracy":0.8,"n":10},"history":{"accuracy":0.5,"n":4},"law":{"accuracy":0.38461538461538464,"n":13},"other":{"accuracy":0.7272727272727273,"n":11},"philosophy":{"accuracy":0.3333333333333333,"n":6},"physics":{"accuracy":0.2,"n":15},"psychology":{"accuracy":1.0,"n":9}},"level":{"":{"accuracy":0.5078125,"n":128}},"n":128,"truncation_rate":0.0}},"teacher":{"math":{"accuracy":0.84375,"category":{"Algebra":{"accuracy":1.0,"n":17},"Counting & Probability":{"accuracy":1.0,"n":7},"Geometry":{"accuracy":0.8333333333333334,"n":6},"Intermediate Algebra":{"accuracy":0.5714285714285714,"n":14},"Number Theory":{"accuracy":0.8333333333333334,"n":6},"Prealgebra":{"accuracy":0.8888888888888888,"n":9},"Precalculus":{"accuracy":0.8,"n":5}},"level":{"1":{"accuracy":1.0,"n":5},"2":{"accuracy":1.0,"n":11},"3":{"accuracy":1.0,"n":15},"4":{"accuracy":0.8125,"n":16},"5":{"accuracy":0.5882352941176471,"n":17}},"n":64,"truncation_rate":0.03125},"retention":{"accuracy":0.140625,"category":{"biology":{"accuracy":0.1111111111111111,"n":9},"business":{"accuracy":0.1,"n":10},"chemistry":{"accuracy":0.14285714285714285,"n":14},"computer science":{"accuracy":0.2,"n":5},"economics":{"accuracy":0.1,"n":10},"engineering":{"accuracy":0.25,"n":12},"health":{"accuracy":0.1,"n":10},"history":{"accuracy":0.0,"n":4},"law":{"accuracy":0.07692307692307693,"n":13},"other":{"accuracy":0.09090909090909091,"n":11},"philosophy":{"accuracy":0.16666666666666666,"n":6},"physics":{"accuracy":0.13333333333333333,"n":15},"psychology":{"accuracy":0.3333333333333333,"n":9}},"level":{"":{"accuracy":0.140625,"n":128}},"n":128,"truncation_rate":0.0}}},"passed":{"target_math_headroom":true,"target_retention_advantage":true,"teacher_math_advantage":true}},"limitations":["One training seed; paired item-bootstrap intervals condition on trained checkpoints.","The selected arm tests a fixed learning-rate scaling heuristic after proxy method selection, not general agent research ability.","Trusted maintainer POC; not a hardened participant-facing benchmark service.","Inference uses 2048 math output tokens and zero-shot MMLU-Pro letter likelihood.","LoRA adaptation is used at both scales; these results do not establish full-parameter transfer.","The verifier uses Numina reference answers; extraction consistency does not establish mathematical label correctness."],"freeze":{"data_hash":"4423b95b487763c4cc9489b421deba6e558bcf1b155328496be66c3cd57cea5c","direct":{"accumulation":4,"alpha":32,"beta":0.02,"group_size":4,"learning_rate":0.0001,"method":"reference_sft","micro_batch":2,"rank":16,"rollout_tokens":1024,"seed":7107,"teacher_fraction":0.0,"temperature":0.8},"frozen_unix":1791523476.5651793,"hypothesis":"Proxy experiment p5-opd is the highest-scoring retention-admissible method under the preregistered rule. Transfer its objective, data mixture, adapter rank, batching and sampling settings. Reduce the target learning rate by sqrt(1.5/14), because the stronger 14B model may need smaller parameter updates to preserve existing reasoning and general capability. This is a preregistered scale heuristic, not an empirically fitted scaling law. The naive arm uses the identical proxy winner without the learning-rate change, isolating that transfer decision. No trained-target development or final scores informed this choice.","naive":{"accumulation":4,"alpha":32,"beta":0.02,"group_size":2,"learning_rate":3e-05,"method":"opd","micro_batch":2,"rank":16,"rollout_tokens":1024,"seed":7107,"teacher_fraction":0.0,"temperature":0.8},"recipe":{"accumulation":4,"alpha":32,"beta":0.02,"group_size":2,"learning_rate":9.819805060619657e-06,"method":"opd","micro_batch":2,"rank":16,"rollout_tokens":1024,"seed":7107,"teacher_fraction":0.0,"temperature":0.8},"sha256":"fa350e5f91744e46b7d739f0df46b45521373c2dbe51cc13d87485bbb834bf98","supporting_proxy_experiments":["p0-reference-mid","p1-reference-high","p2-reference-low","p3-teacher","p4-mixed","p5-opd","p6-grpo-low","p7-grpo-high"]},"study":"f07-math-study-v3-1009","date":"2026-10-09","title":"Cross-scale post-training","models":{"proxy":{"repo":"Qwen/Qwen2.5-1.5B-Instruct","type":"model","revision":"989aa7980e4cf806f80c7fef2b1adb7bc71aa306","license":"apache-2.0"},"target":{"repo":"Qwen/Qwen2.5-14B-Instruct","type":"model","revision":"cf98f3b3bbb457ad9e2bb7baf9a0125b6b88caa8","license":"apache-2.0"},"teacher":{"repo":"Qwen/Qwen2.5-Math-7B-Instruct","type":"model","revision":"ef9926d75ab1d54532f6a30dd5e760355eb9aa4d","license":"apache-2.0"}},"audit":{"frozen_sha256":"fa350e5f91744e46b7d739f0df46b45521373c2dbe51cc13d87485bbb834bf98","method_hashes":{"metrics.py":"7e6d7e2b9e88205f8fb0fe74b3d009f5c210a833d84a0aadd8c02f978ec44996","protocol.py":"71b0b1a398e4d356fda1f2e19c01044446afd023efcb64d359917e56ecbe7aee","runtime.py":"7fb7cbe3599be56719631e4387f4a9c98c4c649f3644c8bf0c115d9f93cba79c","train.py":"abdbd6955594962529ecbb1c72494522f4c2d0a4b60e898a4a72b483aa113303"},"proxy_gpu_seconds":5216.889497671975,"proxy_positions":4186112,"responses_regraded":3536,"scope":"completed trusted maintainer cohort; invalidated debugging costs recorded separately","source_files_verified":1027,"valid":true},"curves":{"p0-reference-mid":{"role":"proxy","recipe":{"accumulation":4,"alpha":32,"beta":0.02,"group_size":4,"learning_rate":0.0001,"method":"reference_sft","micro_batch":2,"rank":16,"rollout_tokens":1024,"seed":7107,"teacher_fraction":0.0,"temperature":0.8},"updates":[{"gradient_norm":0.35889899730682373,"loss":0.6744832172989845,"positions":16384,"seconds":7.568228242918849,"update":1},{"gradient_norm":0.2550978362560272,"loss":0.5336394309997559,"positions":32768,"seconds":8.395871805027127,"update":2},{"gradient_norm":0.22688202559947968,"loss":0.5099123045802116,"positions":49152,"seconds":9.23075434891507,"update":3},{"gradient_norm":0.3003944754600525,"loss":0.6072368323802948,"positions":65536,"seconds":10.037459848914295,"update":4},{"gradient_norm":0.21902471780776978,"loss":0.35271351784467697,"positions":81920,"seconds":10.889772382797673,"update":5},{"gradient_norm":0.17640265822410583,"loss":0.367046058177948,"positions":98304,"seconds":11.699752230895683,"update":6},{"gradient_norm":0.2857263386249542,"loss":0.5579106584191322,"positions":114688,"seconds":12.788811353966594,"update":7},{"gradient_norm":0.17636516690254211,"loss":0.5104538798332214,"positions":131072,"seconds":13.760360118933022,"update":8},{"gradient_norm":0.2057754099369049,"loss":0.46116428077220917,"positions":147456,"seconds":14.583202823996544,"update":9},{"gradient_norm":0.1863013505935669,"loss":0.49889129400253296,"positions":163840,"seconds":15.41977356094867,"update":10},{"gradient_norm":0.18979693949222565,"loss":0.45393499732017517,"positions":180224,"seconds":16.22781771980226,"update":11},{"gradient_norm":0.2766176462173462,"loss":0.6226877570152283,"positions":196608,"seconds":17.016139874933288,"update":12},{"gradient_norm":0.1896323710680008,"loss":0.5045281872153282,"positions":212992,"seconds":17.950382737908512,"update":13},{"gradient_norm":0.2128826230764389,"loss":0.6314587965607643,"positions":229376,"seconds":18.86099939001724,"update":14},{"gradient_norm":0.26460617780685425,"loss":0.49082498252391815,"positions":245760,"seconds":19.724760660901666,"update":15},{"gradient_norm":0.22938422858715057,"loss":0.5561513155698776,"positions":262144,"seconds":20.581973988795653,"update":16},{"gradient_norm":0.2506803572177887,"loss":0.672539621591568,"positions":278528,"seconds":21.44068693788722,"update":17},{"gradient_norm":0.1914617121219635,"loss":0.5118826851248741,"positions":294912,"seconds":22.302523089805618,"update":18},{"gradient_norm":0.2790408730506897,"loss":0.5546391978859901,"positions":311296,"seconds":23.12969416589476,"update":19},{"gradient_norm":0.261056125164032,"loss":0.5080516710877419,"positions":327680,"seconds":24.074801401002333,"update":20},{"gradient_norm":0.24211706221103668,"loss":0.6706522554159164,"positions":344064,"seconds":24.874156422913074,"update":21},{"gradient_norm":0.29213887453079224,"loss":0.5392218828201294,"positions":360448,"seconds":25.694630154874176,"update":22},{"gradient_norm":0.18509432673454285,"loss":0.482999786734581,"positions":376832,"seconds":26.508567353943363,"update":23},{"gradient_norm":0.2619078755378723,"loss":0.6629382446408272,"positions":393216,"seconds":27.307050140807405,"update":24},{"gradient_norm":0.2305060476064682,"loss":0.6130814328789711,"positions":409600,"seconds":28.097909177886322,"update":25},{"gradient_norm":0.22015677392482758,"loss":0.4089370220899582,"positions":425984,"seconds":29.04733033082448,"update":26},{"gradient_norm":0.2067122608423233,"loss":0.5078695937991142,"positions":442368,"seconds":29.82893377589062,"update":27},{"gradient_norm":0.20323258638381958,"loss":0.5464537963271141,"positions":458752,"seconds":30.614957614801824,"update":28},{"gradient_norm":0.19078868627548218,"loss":0.4613889902830124,"positions":475136,"seconds":31.397057439899072,"update":29},{"gradient_norm":0.18961530923843384,"loss":0.41664309054613113,"positions":491520,"seconds":32.17985705193132,"update":30},{"gradient_norm":0.20309282839298248,"loss":0.4310808628797531,"positions":507904,"seconds":32.986528954003006,"update":31},{"gradient_norm":0.21923358738422394,"loss":0.4729415588080883,"positions":524288,"seconds":33.94010375300422,"update":32}],"receipt":{"adapter_sha256":"20b4cce7912903ef757990a6ff48f105baee00bf8c4a73bcc709a20d6a2ce760","peak_allocated_bytes":11305907200,"position_limit":524288,"positions":524288,"recipe":{"accumulation":4,"alpha":32,"beta":0.02,"group_size":4,"learning_rate":0.0001,"method":"reference_sft","micro_batch":2,"rank":16,"rollout_tokens":1024,"seed":7107,"teacher_fraction":0.0,"temperature":0.8},"role":"proxy","teacher_cache_hash":null,"train_seconds":35.04000649298541,"training_data_hash":"1118fbf8abc66267f6e0d05201f62f9d4a9b13ee6a8a9626aa4d04ec03aa1284","updates":32}},"p1-reference-high":{"role":"proxy","recipe":{"accumulation":4,"alpha":32,"beta":0.02,"group_size":4,"learning_rate":0.0003,"method":"reference_sft","micro_batch":2,"rank":16,"rollout_tokens":1024,"seed":7107,"teacher_fraction":0.0,"temperature":0.8},"updates":[{"gradient_norm":0.35898518562316895,"loss":0.6744832172989845,"positions":16384,"seconds":9.527407998917624,"update":1},{"gradient_norm":0.2609086036682129,"loss":0.532670870423317,"positions":32768,"seconds":10.417145910905674,"update":2},{"gradient_norm":0.2185886800289154,"loss":0.50776606798172,"positions":49152,"seconds":11.474166051018983,"update":3},{"gradient_norm":0.2632639408111572,"loss":0.6068605408072472,"positions":65536,"seconds":12.354334677103907,"update":4},{"gradient_norm":0.1902206391096115,"loss":0.34849053621292114,"positions":81920,"seconds":13.223263848107308,"update":5},{"gradient_norm":0.1545008420944214,"loss":0.3633444607257843,"positions":98304,"seconds":14.078713292023167,"update":6},{"gradient_norm":0.2504104673862457,"loss":0.5568336918950081,"positions":114688,"seconds":15.12528808391653,"update":7},{"gradient_norm":0.16815903782844543,"loss":0.5055716261267662,"positions":131072,"seconds":15.980139749124646,"update":8},{"gradient_norm":0.17098650336265564,"loss":0.4568117558956146,"positions":147456,"seconds":16.847637753933668,"update":9},{"gradient_norm":0.16178135573863983,"loss":0.4929546192288399,"positions":163840,"seconds":17.675164239946753,"update":10},{"gradient_norm":0.17983612418174744,"loss":0.4554664120078087,"positions":180224,"seconds":18.52539140312001,"update":11},{"gradient_norm":0.26444491744041443,"loss":0.6223282366991043,"positions":196608,"seconds":19.38738224795088,"update":12},{"gradient_norm":0.18082299828529358,"loss":0.5039727836847305,"positions":212992,"seconds":20.391244559083134,"update":13},{"gradient_norm":0.1947917640209198,"loss":0.628077395260334,"positions":229376,"seconds":21.25230639008805,"update":14},{"gradient_norm":0.2552903890609741,"loss":0.48618247359991074,"positions":245760,"seconds":22.094261445105076,"update":15},{"gradient_norm":0.21945571899414062,"loss":0.5519287884235382,"positions":262144,"seconds":22.942388477968052,"update":16},{"gradient_norm":0.2390042543411255,"loss":0.6748115420341492,"positions":278528,"seconds":23.78538148291409,"update":17},{"gradient_norm":0.1894705593585968,"loss":0.5164574608206749,"positions":294912,"seconds":24.65478425496258,"update":18},{"gradient_norm":0.26677805185317993,"loss":0.5498155653476715,"positions":311296,"seconds":25.49505113810301,"update":19},{"gradient_norm":0.24278150498867035,"loss":0.5041075125336647,"positions":327680,"seconds":26.516936623957008,"update":20},{"gradient_norm":0.2287411093711853,"loss":0.668493889272213,"positions":344064,"seconds":27.35159841994755,"update":21},{"gradient_norm":0.32208484411239624,"loss":0.5476506650447845,"positions":360448,"seconds":28.186171541921794,"update":22},{"gradient_norm":0.17900899052619934,"loss":0.4814933314919472,"positions":376832,"seconds":29.020379419904202,"update":23},{"gradient_norm":0.22341930866241455,"loss":0.6558378264307976,"positions":393216,"seconds":29.864886455005035,"update":24},{"gradient_norm":0.22004547715187073,"loss":0.6107470244169235,"positions":409600,"seconds":30.713065342977643,"update":25},{"gradient_norm":0.2105880230665207,"loss":0.40963253378868103,"positions":425984,"seconds":31.71583663392812,"update":26},{"gradient_norm":0.19546973705291748,"loss":0.5089633241295815,"positions":442368,"seconds":32.57721812510863,"update":27},{"gradient_norm":0.18697300553321838,"loss":0.5459836721420288,"positions":458752,"seconds":33.4283007690683,"update":28},{"gradient_norm":0.1757633537054062,"loss":0.4581832252442837,"positions":475136,"seconds":34.282376520102844,"update":29},{"gradient_norm":0.1864161342382431,"loss":0.41662581264972687,"positions":491520,"seconds":35.13785697403364,"update":30},{"gradient_norm":0.19950975477695465,"loss":0.43761085718870163,"positions":507904,"seconds":35.97095004795119,"update":31},{"gradient_norm":0.204032763838768,"loss":0.4716094806790352,"positions":524288,"seconds":36.94646690809168,"update":32}],"receipt":{"adapter_sha256":"b5293290453cdf4854e9a155cedaec44b503077cbf007cf2a38f604b61e7ea27","peak_allocated_bytes":11305907200,"position_limit":524288,"positions":524288,"recipe":{"accumulation":4,"alpha":32,"beta":0.02,"group_size":4,"learning_rate":0.0003,"method":"reference_sft","micro_batch":2,"rank":16,"rollout_tokens":1024,"seed":7107,"teacher_fraction":0.0,"temperature":0.8},"role":"proxy","teacher_cache_hash":null,"train_seconds":37.87431682390161,"training_data_hash":"1118fbf8abc66267f6e0d05201f62f9d4a9b13ee6a8a9626aa4d04ec03aa1284","updates":32}},"p2-reference-low":{"role":"proxy","recipe":{"accumulation":4,"alpha":32,"beta":0.02,"group_size":4,"learning_rate":3e-05,"method":"reference_sft","micro_batch":2,"rank":16,"rollout_tokens":1024,"seed":7107,"teacher_fraction":0.0,"temperature":0.8},"updates":[{"gradient_norm":0.35898706316947937,"loss":0.6744832172989845,"positions":16384,"seconds":7.927495130104944,"update":1},{"gradient_norm":0.2637330889701843,"loss":0.5357741266489029,"positions":32768,"seconds":8.99807643610984,"update":2},{"gradient_norm":0.2384665310382843,"loss":0.5140655189752579,"positions":49152,"seconds":9.809387376066297,"update":3},{"gradient_norm":0.3395065665245056,"loss":0.6159070581197739,"positions":65536,"seconds":10.638629326131195,"update":4},{"gradient_norm":0.2391304075717926,"loss":0.35860805958509445,"positions":81920,"seconds":11.466628139140084,"update":5},{"gradient_norm":0.19503648579120636,"loss":0.3738902360200882,"positions":98304,"seconds":12.268634079024196,"update":6},{"gradient_norm":0.33293184638023376,"loss":0.5672749876976013,"positions":114688,"seconds":13.22951966803521,"update":7},{"gradient_norm":0.18240003287792206,"loss":0.5166449323296547,"positions":131072,"seconds":14.050470555201173,"update":8},{"gradient_norm":0.22178609669208527,"loss":0.4674346596002579,"positions":147456,"seconds":14.854276088066399,"update":9},{"gradient_norm":0.19371704757213593,"loss":0.5042342618107796,"positions":163840,"seconds":15.669895154191181,"update":10},{"gradient_norm":0.19569920003414154,"loss":0.45668526738882065,"positions":180224,"seconds":16.513591539114714,"update":11},{"gradient_norm":0.30514198541641235,"loss":0.633259229362011,"positions":196608,"seconds":17.34946417203173,"update":12},{"gradient_norm":0.2019248604774475,"loss":0.5134832337498665,"positions":212992,"seconds":18.30338483606465,"update":13},{"gradient_norm":0.22123010456562042,"loss":0.6422203108668327,"positions":229376,"seconds":19.080021672183648,"update":14},{"gradient_norm":0.31581881642341614,"loss":0.5049312338232994,"positions":245760,"seconds":19.870840311050415,"update":15},{"gradient_norm":0.2511937618255615,"loss":0.5669012367725372,"positions":262144,"seconds":20.66746774618514,"update":16},{"gradient_norm":0.2661805748939514,"loss":0.6758534833788872,"positions":278528,"seconds":21.466612498043105,"update":17},{"gradient_norm":0.21780726313591003,"loss":0.5263369232416153,"positions":294912,"seconds":22.25818175612949,"update":18},{"gradient_norm":0.3306991159915924,"loss":0.5687584578990936,"positions":311296,"seconds":23.19419843214564,"update":19},{"gradient_norm":0.283425897359848,"loss":0.515892319381237,"positions":327680,"seconds":23.983699017204344,"update":20},{"gradient_norm":0.2705027461051941,"loss":0.6778868660330772,"positions":344064,"seconds":24.767924485029653,"update":21},{"gradient_norm":0.2896902561187744,"loss":0.5354908406734467,"positions":360448,"seconds":25.55345098511316,"update":22},{"gradient_norm":0.21195870637893677,"loss":0.4884888604283333,"positions":376832,"seconds":26.34372844104655,"update":23},{"gradient_norm":0.29818156361579895,"loss":0.6797882840037346,"positions":393216,"seconds":27.141331700142473,"update":24},{"gradient_norm":0.2566072940826416,"loss":0.6238978430628777,"positions":409600,"seconds":27.959040690213442,"update":25},{"gradient_norm":0.2561652362346649,"loss":0.4142734222114086,"positions":425984,"seconds":28.96817632415332,"update":26},{"gradient_norm":0.22974678874015808,"loss":0.5150314494967461,"positions":442368,"seconds":29.782104283105582,"update":27},{"gradient_norm":0.2251826822757721,"loss":0.555823028087616,"positions":458752,"seconds":30.581040568184108,"update":28},{"gradient_norm":0.21726389229297638,"loss":0.47229764983057976,"positions":475136,"seconds":31.395422634202987,"update":29},{"gradient_norm":0.20861396193504333,"loss":0.41872895136475563,"positions":491520,"seconds":32.21767648705281,"update":30},{"gradient_norm":0.22419624030590057,"loss":0.43248897790908813,"positions":507904,"seconds":32.99637724622153,"update":31},{"gradient_norm":0.2691996991634369,"loss":0.47873474657535553,"positions":524288,"seconds":33.9225508661475,"update":32}],"receipt":{"adapter_sha256":"1e7ff45689cbb4a410dcda9b7457564a800f69f8e485fe9d514be409a6c55159","peak_allocated_bytes":11305907200,"position_limit":524288,"positions":524288,"recipe":{"accumulation":4,"alpha":32,"beta":0.02,"group_size":4,"learning_rate":3e-05,"method":"reference_sft","micro_batch":2,"rank":16,"rollout_tokens":1024,"seed":7107,"teacher_fraction":0.0,"temperature":0.8},"role":"proxy","teacher_cache_hash":null,"train_seconds":34.86071789520793,"training_data_hash":"1118fbf8abc66267f6e0d05201f62f9d4a9b13ee6a8a9626aa4d04ec03aa1284","updates":32}},"p3-teacher":{"role":"proxy","recipe":{"accumulation":4,"alpha":32,"beta":0.02,"group_size":4,"learning_rate":0.0001,"method":"teacher_sft","micro_batch":2,"rank":16,"rollout_tokens":1024,"seed":7107,"teacher_fraction":1.0,"temperature":0.8},"updates":[{"gradient_norm":0.1510794758796692,"loss":0.10750205628573895,"positions":16384,"seconds":11.711338797817007,"update":1},{"gradient_norm":0.15577322244644165,"loss":0.11837903782725334,"positions":32768,"seconds":12.517831455916166,"update":2},{"gradient_norm":0.1825360357761383,"loss":0.13434850983321667,"positions":49152,"seconds":13.360271544894204,"update":3},{"gradient_norm":0.20290037989616394,"loss":0.13780731335282326,"positions":65536,"seconds":14.170799233019352,"update":4},{"gradient_norm":0.13830028474330902,"loss":0.14048493094742298,"positions":81920,"seconds":14.989147607935593,"update":5},{"gradient_norm":0.17324107885360718,"loss":0.14213594049215317,"positions":98304,"seconds":15.787473371019587,"update":6},{"gradient_norm":0.14211341738700867,"loss":0.08037916198372841,"positions":114688,"seconds":16.76472644484602,"update":7},{"gradient_norm":0.18016166985034943,"loss":0.13830736093223095,"positions":131072,"seconds":17.552858717972413,"update":8},{"gradient_norm":0.1653977334499359,"loss":0.11587494425475597,"positions":147456,"seconds":18.34624625882134,"update":9},{"gradient_norm":0.16492034494876862,"loss":0.1061856709420681,"positions":163840,"seconds":19.143334805965424,"update":10},{"gradient_norm":0.15394572913646698,"loss":0.09356828965246677,"positions":180224,"seconds":19.95518137491308,"update":11},{"gradient_norm":0.17827874422073364,"loss":0.09823920577764511,"positions":196608,"seconds":20.787139037856832,"update":12},{"gradient_norm":0.18611523509025574,"loss":0.12197964265942574,"positions":212992,"seconds":21.71735765086487,"update":13},{"gradient_norm":0.18351975083351135,"loss":0.17296796292066574,"positions":229376,"seconds":22.497036849847063,"update":14},{"gradient_norm":0.13211582601070404,"loss":0.09878201689571142,"positions":245760,"seconds":23.284013991011307,"update":15},{"gradient_norm":0.20672036707401276,"loss":0.1880338415503502,"positions":262144,"seconds":24.065684981876984,"update":16},{"gradient_norm":0.18686941266059875,"loss":0.11294577084481716,"positions":278528,"seconds":24.849272336810827,"update":17},{"gradient_norm":0.20970600843429565,"loss":0.12283024564385414,"positions":294912,"seconds":25.62632601382211,"update":18},{"gradient_norm":0.15610823035240173,"loss":0.12122683227062225,"positions":311296,"seconds":26.420191946905106,"update":19},{"gradient_norm":0.1975872665643692,"loss":0.15686251409351826,"positions":327680,"seconds":27.378594042966142,"update":20},{"gradient_norm":0.16674934327602386,"loss":0.10828964598476887,"positions":344064,"seconds":28.19518869696185,"update":21},{"gradient_norm":0.1651816964149475,"loss":0.10220626182854176,"positions":360448,"seconds":28.995108815841377,"update":22},{"gradient_norm":0.17107340693473816,"loss":0.10356495901942253,"positions":376832,"seconds":29.7980002888944,"update":23},{"gradient_norm":0.18272286653518677,"loss":0.08268390502780676,"positions":393216,"seconds":30.60564573900774,"update":24},{"gradient_norm":0.1370772421360016,"loss":0.10651658568531275,"positions":409600,"seconds":31.405317927943543,"update":25},{"gradient_norm":0.17710252106189728,"loss":0.15089867264032364,"positions":425984,"seconds":32.35730983084068,"update":26},{"gradient_norm":0.14606095850467682,"loss":0.11064767092466354,"positions":442368,"seconds":33.14558756700717,"update":27},{"gradient_norm":0.13987967371940613,"loss":0.10008441284298897,"positions":458752,"seconds":33.94313089293428,"update":28},{"gradient_norm":0.17813584208488464,"loss":0.10592105891555548,"positions":475136,"seconds":34.73063310794532,"update":29},{"gradient_norm":0.18400274217128754,"loss":0.11877972073853016,"positions":491520,"seconds":35.676631239941344,"update":30},{"gradient_norm":0.1747414469718933,"loss":0.1381767801940441,"positions":507904,"seconds":36.63683079695329,"update":31},{"gradient_norm":0.1429928094148636,"loss":0.09784555621445179,"positions":524288,"seconds":37.58202443481423,"update":32}],"receipt":{"adapter_sha256":"78759c075f7bcfe06a8185278fae8f996bcc4edda89f93d07181b190ccb9bb2f","peak_allocated_bytes":11305907200,"position_limit":524288,"positions":524288,"recipe":{"accumulation":4,"alpha":32,"beta":0.02,"group_size":4,"learning_rate":0.0001,"method":"teacher_sft","micro_batch":2,"rank":16,"rollout_tokens":1024,"seed":7107,"teacher_fraction":1.0,"temperature":0.8},"role":"proxy","teacher_cache_hash":"6629bea9b6dd9c749d7613305a51f461a8c861be74f96ad4dc072e027759e1ba","train_seconds":38.58267200598493,"training_data_hash":"1118fbf8abc66267f6e0d05201f62f9d4a9b13ee6a8a9626aa4d04ec03aa1284","updates":32}},"p4-mixed":{"role":"proxy","recipe":{"accumulation":4,"alpha":32,"beta":0.02,"group_size":4,"learning_rate":0.0001,"method":"mixed_sft","micro_batch":2,"rank":16,"rollout_tokens":1024,"seed":7107,"teacher_fraction":0.5,"temperature":0.8},"updates":[{"gradient_norm":0.2630555331707001,"loss":0.19009336829185486,"positions":16384,"seconds":8.351020700996742,"update":1},{"gradient_norm":0.2700391709804535,"loss":0.5219943411648273,"positions":32768,"seconds":9.294762591132894,"update":2},{"gradient_norm":0.1689787060022354,"loss":0.286625150591135,"positions":49152,"seconds":10.228692305041477,"update":3},{"gradient_norm":0.2170805186033249,"loss":0.49298304319381714,"positions":65536,"seconds":11.151529202936217,"update":4},{"gradient_norm":0.22202423214912415,"loss":0.2919960804283619,"positions":81920,"seconds":12.088497777003795,"update":5},{"gradient_norm":0.1785278618335724,"loss":0.29746175184845924,"positions":98304,"seconds":12.991593247978017,"update":6},{"gradient_norm":0.18768514692783356,"loss":0.48603080213069916,"positions":114688,"seconds":14.096260681049898,"update":7},{"gradient_norm":0.157619908452034,"loss":0.14453650638461113,"positions":131072,"seconds":15.019584933063015,"update":8},{"gradient_norm":0.28862088918685913,"loss":0.32210076227784157,"positions":147456,"seconds":15.941337265074253,"update":9},{"gradient_norm":0.2630385458469391,"loss":0.2806067653000355,"positions":163840,"seconds":16.855158258927986,"update":10},{"gradient_norm":0.1753566414117813,"loss":0.30637577548623085,"positions":180224,"seconds":17.79605496698059,"update":11},{"gradient_norm":0.14770659804344177,"loss":0.1669324804097414,"positions":196608,"seconds":18.72710682800971,"update":12},{"gradient_norm":0.16553384065628052,"loss":0.1726763602346182,"positions":212992,"seconds":19.811632097000256,"update":13},{"gradient_norm":0.3100802004337311,"loss":0.4890645071864128,"positions":229376,"seconds":20.711765780113637,"update":14},{"gradient_norm":0.15146739780902863,"loss":0.20466071739792824,"positions":245760,"seconds":21.63095709704794,"update":15},{"gradient_norm":0.14616675674915314,"loss":0.29055233485996723,"positions":262144,"seconds":22.539752956014127,"update":16},{"gradient_norm":0.1811390370130539,"loss":0.29612886905670166,"positions":278528,"seconds":23.453913626028225,"update":17},{"gradient_norm":0.15091785788536072,"loss":0.20309916138648987,"positions":294912,"seconds":24.36626468691975,"update":18},{"gradient_norm":0.1533547192811966,"loss":0.3414524607360363,"positions":311296,"seconds":25.291334904031828,"update":19},{"gradient_norm":0.16366145014762878,"loss":0.2059057541191578,"positions":327680,"seconds":26.353749515954405,"update":20},{"gradient_norm":0.26213958859443665,"loss":0.28515099734067917,"positions":344064,"seconds":27.26111258403398,"update":21},{"gradient_norm":0.17245711386203766,"loss":0.32460034638643265,"positions":360448,"seconds":28.16273715905845,"update":22},{"gradient_norm":0.19782321155071259,"loss":0.3926243185997009,"positions":376832,"seconds":29.074153319001198,"update":23},{"gradient_norm":0.31681469082832336,"loss":0.6293707862496376,"positions":393216,"seconds":29.993294012965634,"update":24},{"gradient_norm":0.17346566915512085,"loss":0.2794817164540291,"positions":409600,"seconds":30.90737779601477,"update":25},{"gradient_norm":0.18646833300590515,"loss":0.31030495930463076,"positions":425984,"seconds":32.00955880712718,"update":26},{"gradient_norm":0.13938894867897034,"loss":0.23571731708943844,"positions":442368,"seconds":32.92167395702563,"update":27},{"gradient_norm":0.19433696568012238,"loss":0.4671209864318371,"positions":458752,"seconds":33.84965126612224,"update":28},{"gradient_norm":0.28696221113204956,"loss":0.42474014312028885,"positions":475136,"seconds":34.789057257119566,"update":29},{"gradient_norm":0.1569332331418991,"loss":0.1381067242473364,"positions":491520,"seconds":35.70915669412352,"update":30},{"gradient_norm":0.1816251128911972,"loss":0.28751440346241,"positions":507904,"seconds":36.61268842197023,"update":31},{"gradient_norm":0.2898440659046173,"loss":0.49132461100816727,"positions":524288,"seconds":37.50959458993748,"update":32}],"receipt":{"adapter_sha256":"5b4db0af97bd2a754fb52bd2363f0702ec444d4dd30a8ec55f23a0e7ad2bfd0d","peak_allocated_bytes":11305907200,"position_limit":524288,"positions":524288,"recipe":{"accumulation":4,"alpha":32,"beta":0.02,"group_size":4,"learning_rate":0.0001,"method":"mixed_sft","micro_batch":2,"rank":16,"rollout_tokens":1024,"seed":7107,"teacher_fraction":0.5,"temperature":0.8},"role":"proxy","teacher_cache_hash":"6629bea9b6dd9c749d7613305a51f461a8c861be74f96ad4dc072e027759e1ba","train_seconds":38.62453570798971,"training_data_hash":"1118fbf8abc66267f6e0d05201f62f9d4a9b13ee6a8a9626aa4d04ec03aa1284","updates":32}},"p5-opd":{"role":"proxy","recipe":{"accumulation":4,"alpha":32,"beta":0.02,"group_size":2,"learning_rate":3e-05,"method":"opd","micro_batch":2,"rank":16,"rollout_tokens":1024,"seed":7107,"teacher_fraction":0.0,"temperature":0.8},"updates":[{"gradient_norm":3.9525146484375,"loss":-8.89499282836914,"mean_reward":null,"positions":12288,"seconds":57.13853821600787,"update":1},{"gradient_norm":3.3195064067840576,"loss":-3.2220988273620605,"mean_reward":null,"positions":24576,"seconds":77.04313017404638,"update":2},{"gradient_norm":2.8920044898986816,"loss":0.3994755744934082,"mean_reward":null,"positions":36864,"seconds":125.42416800395586,"update":3},{"gradient_norm":2.340399742126465,"loss":-2.5831246376037598,"mean_reward":null,"positions":49152,"seconds":157.90920456196181,"update":4},{"gradient_norm":4.102895736694336,"loss":1.058940052986145,"mean_reward":null,"positions":61440,"seconds":181.79241067101248,"update":5},{"gradient_norm":2.099170446395874,"loss":0.2017340362071991,"mean_reward":null,"positions":73728,"seconds":216.47727214405313,"update":6},{"gradient_norm":2.3957958221435547,"loss":0.5283023715019226,"mean_reward":null,"positions":86016,"seconds":241.9015394288581,"update":7},{"gradient_norm":2.6537563800811768,"loss":1.3011059761047363,"mean_reward":null,"positions":98304,"seconds":267.464616863057,"update":8},{"gradient_norm":3.1331734657287598,"loss":-6.475659370422363,"mean_reward":null,"positions":110592,"seconds":288.7540947289672,"update":9},{"gradient_norm":3.01781964302063,"loss":0.15658503770828247,"mean_reward":null,"positions":122880,"seconds":299.651006228989,"update":10},{"gradient_norm":3.4535574913024902,"loss":-0.37601345777511597,"mean_reward":null,"positions":135168,"seconds":324.86729261488654,"update":11},{"gradient_norm":3.6235761642456055,"loss":-7.3874030113220215,"mean_reward":null,"positions":147456,"seconds":346.3303809198551,"update":12},{"gradient_norm":3.809359312057495,"loss":-1.8896855115890503,"mean_reward":null,"positions":159744,"seconds":365.6945594889112,"update":13},{"gradient_norm":4.054019451141357,"loss":-6.190342903137207,"mean_reward":null,"positions":172032,"seconds":384.2083551778924,"update":14},{"gradient_norm":2.4850351810455322,"loss":-2.346013307571411,"mean_reward":null,"positions":184320,"seconds":415.94689444685355,"update":15},{"gradient_norm":4.06138277053833,"loss":-1.287091851234436,"mean_reward":null,"positions":196608,"seconds":426.8341421559453,"update":16},{"gradient_norm":4.449258327484131,"loss":-1.7085012197494507,"mean_reward":null,"positions":208896,"seconds":442.1346003478393,"update":17},{"gradient_norm":3.819042205810547,"loss":0.20490938425064087,"mean_reward":null,"positions":221184,"seconds":465.91463375487365,"update":18},{"gradient_norm":4.915318489074707,"loss":-6.805636405944824,"mean_reward":null,"positions":233472,"seconds":482.0381392510608,"update":19},{"gradient_norm":2.067619800567627,"loss":0.5965104699134827,"mean_reward":null,"positions":245760,"seconds":498.65475633298047,"update":20},{"gradient_norm":3.4565298557281494,"loss":-8.246166229248047,"mean_reward":null,"positions":258048,"seconds":525.9981983369216,"update":21},{"gradient_norm":2.5516138076782227,"loss":-0.7513110637664795,"mean_reward":null,"positions":270336,"seconds":561.1553634318989,"update":22},{"gradient_norm":2.1031851768493652,"loss":-0.06554122269153595,"mean_reward":null,"positions":282624,"seconds":579.1431637599599,"update":23},{"gradient_norm":3.2249932289123535,"loss":-0.9944154620170593,"mean_reward":null,"positions":294912,"seconds":609.94479631586,"update":24},{"gradient_norm":3.08874773979187,"loss":1.016690969467163,"mean_reward":null,"positions":307200,"seconds":619.7972191160079,"update":25},{"gradient_norm":3.6254775524139404,"loss":-4.589458465576172,"mean_reward":null,"positions":319488,"seconds":637.341200487921,"update":26},{"gradient_norm":3.5033507347106934,"loss":0.4817572832107544,"mean_reward":null,"positions":331776,"seconds":651.3360679689795,"update":27},{"gradient_norm":3.572756290435791,"loss":-1.3078176975250244,"mean_reward":null,"positions":344064,"seconds":678.2037513840478,"update":28},{"gradient_norm":3.975830316543579,"loss":-4.093016147613525,"mean_reward":null,"positions":356352,"seconds":695.5422898069955,"update":29},{"gradient_norm":4.578216075897217,"loss":1.098093032836914,"mean_reward":null,"positions":368640,"seconds":717.0148453940637,"update":30},{"gradient_norm":1.7319107055664062,"loss":-0.30255407094955444,"mean_reward":null,"positions":380928,"seconds":742.0305321749765,"update":31},{"gradient_norm":3.562213897705078,"loss":-5.038041114807129,"mean_reward":null,"positions":393216,"seconds":764.9933589650318,"update":32},{"gradient_norm":5.0858049392700195,"loss":-0.25306200981140137,"mean_reward":null,"positions":405504,"seconds":779.5481319390237,"update":33},{"gradient_norm":2.9448294639587402,"loss":-3.459608316421509,"mean_reward":null,"positions":417792,"seconds":801.0064249499701,"update":34},{"gradient_norm":2.6396093368530273,"loss":-0.7264207601547241,"mean_reward":null,"positions":430080,"seconds":819.6100418379065,"update":35},{"gradient_norm":2.4670350551605225,"loss":-1.3357816934585571,"mean_reward":null,"positions":442368,"seconds":844.950846910011,"update":36},{"gradient_norm":3.1903977394104004,"loss":-1.7490447759628296,"mean_reward":null,"positions":454656,"seconds":873.0900272929575,"update":37},{"gradient_norm":2.631070137023926,"loss":-2.8589236736297607,"mean_reward":null,"positions":466944,"seconds":890.5351528120227,"update":38},{"gradient_norm":2.962498664855957,"loss":-6.782855033874512,"mean_reward":null,"positions":479232,"seconds":925.0346673650201,"update":39},{"gradient_norm":3.9499435424804688,"loss":-3.442208766937256,"mean_reward":null,"positions":491520,"seconds":948.6981947629247,"update":40},{"gradient_norm":6.4496893882751465,"loss":-3.155365467071533,"mean_reward":null,"positions":503808,"seconds":964.0759486858733,"update":41},{"gradient_norm":4.093798637390137,"loss":-1.5553747415542603,"mean_reward":null,"positions":516096,"seconds":978.2748199899215,"update":42}],"receipt":{"adapter_sha256":"1369a5be794086cc62714d17c689e1b6293b5cfdb30426c3eeb4a1844d5906aa","peak_allocated_bytes":23784590336,"position_limit":524288,"positions":516096,"recipe":{"accumulation":4,"alpha":32,"beta":0.02,"group_size":2,"learning_rate":3e-05,"method":"opd","micro_batch":2,"rank":16,"rollout_tokens":1024,"seed":7107,"teacher_fraction":0.0,"temperature":0.8},"role":"proxy","teacher_cache_hash":null,"train_seconds":979.4048202189151,"training_data_hash":"1118fbf8abc66267f6e0d05201f62f9d4a9b13ee6a8a9626aa4d04ec03aa1284","updates":42}},"p6-grpo-low":{"role":"proxy","recipe":{"accumulation":4,"alpha":32,"beta":0.02,"group_size":4,"learning_rate":3e-05,"method":"grpo","micro_batch":2,"rank":16,"rollout_tokens":1024,"seed":7107,"teacher_fraction":0.0,"temperature":0.8},"updates":[{"gradient_norm":0.0,"loss":0.0,"mean_reward":0.0,"positions":32768,"seconds":56.877526668831706,"update":1},{"gradient_norm":0.5108131766319275,"loss":0.0,"mean_reward":0.25,"positions":65536,"seconds":75.11167394602671,"update":2},{"gradient_norm":0.3736235201358795,"loss":1.4036893844604492e-05,"mean_reward":0.75,"positions":98304,"seconds":94.57116247992963,"update":3},{"gradient_norm":0.0012902485905215144,"loss":2.34266171901254e-05,"mean_reward":0.0,"positions":131072,"seconds":127.02464947081171,"update":4},{"gradient_norm":0.5201601386070251,"loss":2.422928810119629e-05,"mean_reward":0.5,"positions":163840,"seconds":148.71677463199012,"update":5},{"gradient_norm":0.0012443852610886097,"loss":1.8016240574070252e-05,"mean_reward":0.0,"positions":196608,"seconds":184.95446442696266,"update":6},{"gradient_norm":0.4544270932674408,"loss":2.0951032638549805e-05,"mean_reward":0.75,"positions":229376,"seconds":203.5456097018905,"update":7},{"gradient_norm":0.0012515070848166943,"loss":1.6076119209174067e-05,"mean_reward":1.0,"positions":262144,"seconds":226.50803583394736,"update":8},{"gradient_norm":0.0017008575377985835,"loss":2.7253179723629728e-05,"mean_reward":0.0,"positions":294912,"seconds":257.7611936549656,"update":9},{"gradient_norm":0.4542379379272461,"loss":2.3692846298217773e-05,"mean_reward":0.25,"positions":327680,"seconds":283.7397698548157,"update":10},{"gradient_norm":0.0019021949265152216,"loss":3.559313336154446e-05,"mean_reward":0.0,"positions":360448,"seconds":309.4889326079283,"update":11},{"gradient_norm":0.0014755175216123462,"loss":3.54038093064446e-05,"mean_reward":0.0,"positions":393216,"seconds":328.42587088397704,"update":12},{"gradient_norm":0.4774692952632904,"loss":5.1975250244140625e-05,"mean_reward":0.25,"positions":425984,"seconds":350.7985007718671,"update":13},{"gradient_norm":0.0020621910225600004,"loss":2.8128260964876972e-05,"mean_reward":0.0,"positions":458752,"seconds":371.58629612484947,"update":14},{"gradient_norm":0.0021966698113828897,"loss":4.1485945985186845e-05,"mean_reward":0.0,"positions":491520,"seconds":408.34033913887106,"update":15},{"gradient_norm":0.0050413478165864944,"loss":6.141368066892028e-05,"mean_reward":0.0,"positions":524288,"seconds":417.89339382201433,"update":16}],"receipt":{"adapter_sha256":"d642e834883e9d6eda29b494a9ae10b3d1e27dc734b37250a21d1346aec74527","peak_allocated_bytes":13638536704,"position_limit":524288,"positions":524288,"recipe":{"accumulation":4,"alpha":32,"beta":0.02,"group_size":4,"learning_rate":3e-05,"method":"grpo","micro_batch":2,"rank":16,"rollout_tokens":1024,"seed":7107,"teacher_fraction":0.0,"temperature":0.8},"role":"proxy","teacher_cache_hash":null,"train_seconds":419.01613774499856,"training_data_hash":"1118fbf8abc66267f6e0d05201f62f9d4a9b13ee6a8a9626aa4d04ec03aa1284","updates":16}},"p7-grpo-high":{"role":"proxy","recipe":{"accumulation":4,"alpha":32,"beta":0.02,"group_size":4,"learning_rate":0.0001,"method":"grpo","micro_batch":2,"rank":16,"rollout_tokens":1024,"seed":7107,"teacher_fraction":0.0,"temperature":0.8},"updates":[{"gradient_norm":0.0,"loss":0.0,"mean_reward":0.0,"positions":32768,"seconds":49.734646540135145,"update":1},{"gradient_norm":0.5107827186584473,"loss":0.0,"mean_reward":0.25,"positions":65536,"seconds":65.18135980702937,"update":2},{"gradient_norm":0.4004174470901489,"loss":1.8656253814697266e-05,"mean_reward":0.75,"positions":98304,"seconds":81.62989695300348,"update":3},{"gradient_norm":0.0037494273856282234,"loss":6.373053474817425e-05,"mean_reward":0.0,"positions":131072,"seconds":115.22331465920433,"update":4},{"gradient_norm":0.4623227119445801,"loss":0.00014728307723999023,"mean_reward":0.25,"positions":163840,"seconds":130.43393980711699,"update":5},{"gradient_norm":0.003432739060372114,"loss":0.00011391336738597602,"mean_reward":0.0,"positions":196608,"seconds":162.17399080400355,"update":6},{"gradient_norm":0.5669704675674438,"loss":8.249282836914062e-05,"mean_reward":0.25,"positions":229376,"seconds":175.43615353200585,"update":7},{"gradient_norm":0.009711228311061859,"loss":0.0002519300614949316,"mean_reward":1.0,"positions":262144,"seconds":194.03805420920253,"update":8},{"gradient_norm":0.004274844657629728,"loss":0.00012537172005977482,"mean_reward":0.0,"positions":294912,"seconds":213.20910921506584,"update":9},{"gradient_norm":0.04602602869272232,"loss":0.0010287822224199772,"mean_reward":0.0,"positions":327680,"seconds":227.94554989598691,"update":10},{"gradient_norm":0.01245764922350645,"loss":0.000490012695081532,"mean_reward":0.0,"positions":360448,"seconds":245.47071307105944,"update":11},{"gradient_norm":0.006218597758561373,"loss":0.0002481286064721644,"mean_reward":0.0,"positions":393216,"seconds":266.9354016110301,"update":12},{"gradient_norm":0.4783290922641754,"loss":0.0007420182228088379,"mean_reward":0.25,"positions":425984,"seconds":286.4919690650422,"update":13},{"gradient_norm":0.48064032196998596,"loss":0.0008746832609176636,"mean_reward":0.5,"positions":458752,"seconds":305.1663064341992,"update":14},{"gradient_norm":0.01512373797595501,"loss":0.0006088601658120751,"mean_reward":0.0,"positions":491520,"seconds":338.0477071381174,"update":15},{"gradient_norm":0.008833603002130985,"loss":0.00024265200772788376,"mean_reward":0.0,"positions":524288,"seconds":346.7812798351515,"update":16}],"receipt":{"adapter_sha256":"47103ee68b1136eec763025d18701babd46b2a23c8b218f32ea228ceadfbc756","peak_allocated_bytes":13638536704,"position_limit":524288,"positions":524288,"recipe":{"accumulation":4,"alpha":32,"beta":0.02,"group_size":4,"learning_rate":0.0001,"method":"grpo","micro_batch":2,"rank":16,"rollout_tokens":1024,"seed":7107,"teacher_fraction":0.0,"temperature":0.8},"role":"proxy","teacher_cache_hash":null,"train_seconds":348.0023534731008,"training_data_hash":"1118fbf8abc66267f6e0d05201f62f9d4a9b13ee6a8a9626aa4d04ec03aa1284","updates":16}},"direct":{"role":"target","recipe":{"accumulation":4,"alpha":32,"beta":0.02,"group_size":4,"learning_rate":0.0001,"method":"reference_sft","micro_batch":2,"rank":16,"rollout_tokens":1024,"seed":7107,"teacher_fraction":0.0,"temperature":0.8},"updates":[{"gradient_norm":0.36490970849990845,"loss":0.6512189954519272,"positions":16384,"seconds":17.693087412044406,"update":1},{"gradient_norm":0.22836089134216309,"loss":0.500582292675972,"positions":32768,"seconds":20.484491399023682,"update":2},{"gradient_norm":0.22272883355617523,"loss":0.4656769558787346,"positions":49152,"seconds":23.378091000951827,"update":3},{"gradient_norm":0.3604913353919983,"loss":0.5367085821926594,"positions":65536,"seconds":26.15903507382609,"update":4},{"gradient_norm":0.1934966892004013,"loss":0.33259502053260803,"positions":81920,"seconds":28.940322641981766,"update":5},{"gradient_norm":0.12994353473186493,"loss":0.3265910744667053,"positions":98304,"seconds":31.830301705980673,"update":6},{"gradient_norm":0.21893447637557983,"loss":0.5030923932790756,"positions":114688,"seconds":34.596258010016754,"update":7},{"gradient_norm":0.13830657303333282,"loss":0.4346468821167946,"positions":131072,"seconds":37.37196726491675,"update":8},{"gradient_norm":0.18184877932071686,"loss":0.4139245077967644,"positions":147456,"seconds":40.222945685032755,"update":9},{"gradient_norm":0.16762536764144897,"loss":0.44020991027355194,"positions":163840,"seconds":42.999540503835306,"update":10},{"gradient_norm":0.12481648474931717,"loss":0.37516458332538605,"positions":180224,"seconds":45.76676401402801,"update":11},{"gradient_norm":0.21106866002082825,"loss":0.5426698960363865,"positions":196608,"seconds":48.6146371930372,"update":12},{"gradient_norm":0.140332892537117,"loss":0.43069517612457275,"positions":212992,"seconds":51.380200100829825,"update":13},{"gradient_norm":0.16801585257053375,"loss":0.5477978810667992,"positions":229376,"seconds":54.14740735688247,"update":14},{"gradient_norm":0.1608092486858368,"loss":0.41798027232289314,"positions":245760,"seconds":57.029207499930635,"update":15},{"gradient_norm":0.14772264659404755,"loss":0.47787535190582275,"positions":262144,"seconds":59.799738032976165,"update":16},{"gradient_norm":0.16356860101222992,"loss":0.5323083475232124,"positions":278528,"seconds":62.558910246007144,"update":17},{"gradient_norm":0.10904759913682938,"loss":0.4508426859974861,"positions":294912,"seconds":65.41705503198318,"update":18},{"gradient_norm":0.18543218076229095,"loss":0.4583415687084198,"positions":311296,"seconds":68.18083697604015,"update":19},{"gradient_norm":0.1523934155702591,"loss":0.4075492173433304,"positions":327680,"seconds":70.94993667304516,"update":20},{"gradient_norm":0.16012096405029297,"loss":0.5461755990982056,"positions":344064,"seconds":73.73054142692126,"update":21},{"gradient_norm":0.2075078934431076,"loss":0.4638199433684349,"positions":360448,"seconds":76.59087798302062,"update":22},{"gradient_norm":0.1468840390443802,"loss":0.4046652913093567,"positions":376832,"seconds":79.36268217000179,"update":23},{"gradient_norm":0.15068532526493073,"loss":0.5779100432991982,"positions":393216,"seconds":82.13589353393763,"update":24},{"gradient_norm":0.1455351710319519,"loss":0.5360875204205513,"positions":409600,"seconds":85.06870821886696,"update":25},{"gradient_norm":0.15158790349960327,"loss":0.3520028479397297,"positions":425984,"seconds":87.84866825002246,"update":26},{"gradient_norm":0.14383235573768616,"loss":0.434062197804451,"positions":442368,"seconds":90.62042098585516,"update":27},{"gradient_norm":0.13671471178531647,"loss":0.4822896271944046,"positions":458752,"seconds":93.48559882584959,"update":28},{"gradient_norm":0.11305943131446838,"loss":0.3992314897477627,"positions":475136,"seconds":96.25316851888783,"update":29},{"gradient_norm":0.10411544889211655,"loss":0.3624073602259159,"positions":491520,"seconds":99.02124756597914,"update":30},{"gradient_norm":0.12932398915290833,"loss":0.35333408787846565,"positions":507904,"seconds":101.89024171093479,"update":31},{"gradient_norm":0.14626972377300262,"loss":0.42135705798864365,"positions":524288,"seconds":104.67317859688774,"update":32}],"receipt":{"adapter_sha256":"81a13e25154c276dd2ac42b9e60ccad4566d907862ae0434bb6b8c6bb47a6524","peak_allocated_bytes":40442245120,"position_limit":524288,"positions":524288,"recipe":{"accumulation":4,"alpha":32,"beta":0.02,"group_size":4,"learning_rate":0.0001,"method":"reference_sft","micro_batch":2,"rank":16,"rollout_tokens":1024,"seed":7107,"teacher_fraction":0.0,"temperature":0.8},"role":"target","teacher_cache_hash":null,"train_seconds":107.60353357088752,"training_data_hash":"1118fbf8abc66267f6e0d05201f62f9d4a9b13ee6a8a9626aa4d04ec03aa1284","updates":32}},"naive":{"role":"target","recipe":{"accumulation":4,"alpha":32,"beta":0.02,"group_size":2,"learning_rate":3e-05,"method":"opd","micro_batch":2,"rank":16,"rollout_tokens":1024,"seed":7107,"teacher_fraction":0.0,"temperature":0.8},"updates":[{"gradient_norm":0.8826537132263184,"loss":-1.0481349229812622,"mean_reward":null,"positions":12288,"seconds":95.86233712895773,"update":1},{"gradient_norm":0.41093865036964417,"loss":-0.03981976956129074,"mean_reward":null,"positions":24576,"seconds":123.01717453193851,"update":2},{"gradient_norm":0.648910641670227,"loss":-0.2588100731372833,"mean_reward":null,"positions":36864,"seconds":157.24045494990423,"update":3},{"gradient_norm":0.505579948425293,"loss":-0.2046540379524231,"mean_reward":null,"positions":49152,"seconds":214.9588772780262,"update":4},{"gradient_norm":0.6278011202812195,"loss":-0.13973109424114227,"mean_reward":null,"positions":61440,"seconds":241.09813123196363,"update":5},{"gradient_norm":0.2760976552963257,"loss":-0.07724970579147339,"mean_reward":null,"positions":73728,"seconds":297.86262653488666,"update":6},{"gradient_norm":0.5548496842384338,"loss":-0.056242819875478745,"mean_reward":null,"positions":86016,"seconds":323.61525006894954,"update":7},{"gradient_norm":0.17378197610378265,"loss":-0.03580060601234436,"mean_reward":null,"positions":98304,"seconds":360.61759885703214,"update":8},{"gradient_norm":0.33407217264175415,"loss":-0.03978225588798523,"mean_reward":null,"positions":110592,"seconds":412.5952252189163,"update":9},{"gradient_norm":0.619737446308136,"loss":-0.09936045110225677,"mean_reward":null,"positions":122880,"seconds":440.88992391107604,"update":10},{"gradient_norm":0.6016890406608582,"loss":-0.08896224200725555,"mean_reward":null,"positions":135168,"seconds":472.73072987399064,"update":11},{"gradient_norm":0.9100667834281921,"loss":-0.266701340675354,"mean_reward":null,"positions":147456,"seconds":512.7504628789611,"update":12},{"gradient_norm":0.16192325949668884,"loss":-0.0029235382098704576,"mean_reward":null,"positions":159744,"seconds":553.5688879939262,"update":13},{"gradient_norm":0.42625153064727783,"loss":-0.027927391231060028,"mean_reward":null,"positions":172032,"seconds":598.4857384481002,"update":14},{"gradient_norm":0.2555924355983734,"loss":-0.04223480448126793,"mean_reward":null,"positions":184320,"seconds":657.2726878589019,"update":15},{"gradient_norm":2.1440649032592773,"loss":-0.32682090997695923,"mean_reward":null,"positions":196608,"seconds":675.2595946309157,"update":16},{"gradient_norm":1.7119452953338623,"loss":-0.5111455917358398,"mean_reward":null,"positions":208896,"seconds":712.041837405879,"update":17},{"gradient_norm":0.5895478129386902,"loss":-0.05549143627285957,"mean_reward":null,"positions":221184,"seconds":745.8168307119049,"update":18},{"gradient_norm":0.45831069350242615,"loss":-0.05778246372938156,"mean_reward":null,"positions":233472,"seconds":797.4797621620819,"update":19},{"gradient_norm":0.6508468389511108,"loss":-0.024892505258321762,"mean_reward":null,"positions":245760,"seconds":818.7240087499376,"update":20},{"gradient_norm":0.7384347319602966,"loss":-0.04271705448627472,"mean_reward":null,"positions":258048,"seconds":844.4831741170492,"update":21},{"gradient_norm":0.28263744711875916,"loss":-0.009558564983308315,"mean_reward":null,"positions":270336,"seconds":903.143772674026,"update":22},{"gradient_norm":0.7809494733810425,"loss":-0.01685217022895813,"mean_reward":null,"positions":282624,"seconds":929.8484456199221,"update":23},{"gradient_norm":0.41449224948883057,"loss":-0.02033265493810177,"mean_reward":null,"positions":294912,"seconds":984.627623351058,"update":24},{"gradient_norm":0.35784390568733215,"loss":-0.023102151229977608,"mean_reward":null,"positions":307200,"seconds":1002.2814996570814,"update":25},{"gradient_norm":1.3983237743377686,"loss":-0.1444065421819687,"mean_reward":null,"positions":319488,"seconds":1043.8718746600207,"update":26},{"gradient_norm":0.43353480100631714,"loss":0.009820093400776386,"mean_reward":null,"positions":331776,"seconds":1064.6885971189477,"update":27},{"gradient_norm":0.7077815532684326,"loss":-0.09754633158445358,"mean_reward":null,"positions":344064,"seconds":1097.190885660937,"update":28},{"gradient_norm":0.47837063670158386,"loss":-0.07181204855442047,"mean_reward":null,"positions":356352,"seconds":1154.0366481379606,"update":29},{"gradient_norm":0.39239826798439026,"loss":-0.01061995979398489,"mean_reward":null,"positions":368640,"seconds":1201.3295219528954,"update":30},{"gradient_norm":0.649995744228363,"loss":-0.09371200203895569,"mean_reward":null,"positions":380928,"seconds":1246.8688885609154,"update":31},{"gradient_norm":0.698659360408783,"loss":-0.16727319359779358,"mean_reward":null,"positions":393216,"seconds":1303.7331922899466,"update":32},{"gradient_norm":0.9112367033958435,"loss":-0.021023280918598175,"mean_reward":null,"positions":405504,"seconds":1323.7898997259326,"update":33},{"gradient_norm":0.6863943934440613,"loss":-0.07581937313079834,"mean_reward":null,"positions":417792,"seconds":1368.7946547949687,"update":34},{"gradient_norm":0.4572201669216156,"loss":-0.0545317642390728,"mean_reward":null,"positions":430080,"seconds":1400.2442946908996,"update":35},{"gradient_norm":0.415801465511322,"loss":-0.047967538237571716,"mean_reward":null,"positions":442368,"seconds":1455.7650175029412,"update":36},{"gradient_norm":0.5024044513702393,"loss":0.02203366532921791,"mean_reward":null,"positions":454656,"seconds":1492.0147067899816,"update":37},{"gradient_norm":0.7720597982406616,"loss":-0.1360929012298584,"mean_reward":null,"positions":466944,"seconds":1547.719485502923,"update":38},{"gradient_norm":0.3084098696708679,"loss":-0.042849428951740265,"mean_reward":null,"positions":479232,"seconds":1602.609295248054,"update":39},{"gradient_norm":0.4779668152332306,"loss":0.01236659474670887,"mean_reward":null,"positions":491520,"seconds":1629.5913405809551,"update":40},{"gradient_norm":0.6939533352851868,"loss":0.011721758171916008,"mean_reward":null,"positions":503808,"seconds":1655.1434798180126,"update":41},{"gradient_norm":1.1448209285736084,"loss":0.014427149668335915,"mean_reward":null,"positions":516096,"seconds":1681.6602058729623,"update":42}],"receipt":{"adapter_sha256":"c963b45890ecb83fb8c730bd2d2ef29517aef6ff32e0431ac5a097796c8fcc65","peak_allocated_bytes":52718982144,"position_limit":524288,"positions":516096,"recipe":{"accumulation":4,"alpha":32,"beta":0.02,"group_size":2,"learning_rate":3e-05,"method":"opd","micro_batch":2,"rank":16,"rollout_tokens":1024,"seed":7107,"teacher_fraction":0.0,"temperature":0.8},"role":"target","teacher_cache_hash":null,"train_seconds":1684.7486601730343,"training_data_hash":"1118fbf8abc66267f6e0d05201f62f9d4a9b13ee6a8a9626aa4d04ec03aa1284","updates":42}},"selected":{"role":"target","recipe":{"accumulation":4,"alpha":32,"beta":0.02,"group_size":2,"learning_rate":9.819805060619657e-06,"method":"opd","micro_batch":2,"rank":16,"rollout_tokens":1024,"seed":7107,"teacher_fraction":0.0,"temperature":0.8},"updates":[{"gradient_norm":0.8827235698699951,"loss":-1.0481349229812622,"mean_reward":null,"positions":12288,"seconds":97.13548263395205,"update":1},{"gradient_norm":0.4677666425704956,"loss":-0.07886938750743866,"mean_reward":null,"positions":24576,"seconds":118.89417522982694,"update":2},{"gradient_norm":0.5770560503005981,"loss":-0.23240146040916443,"mean_reward":null,"positions":36864,"seconds":154.34250070294365,"update":3},{"gradient_norm":0.3587806224822998,"loss":-0.1347692608833313,"mean_reward":null,"positions":49152,"seconds":217.25503087998368,"update":4},{"gradient_norm":0.5062041878700256,"loss":-0.07728312909603119,"mean_reward":null,"positions":61440,"seconds":243.53153618285432,"update":5},{"gradient_norm":0.299816370010376,"loss":-0.07355111092329025,"mean_reward":null,"positions":73728,"seconds":304.3983656249475,"update":6},{"gradient_norm":0.8540717363357544,"loss":-0.14828111231327057,"mean_reward":null,"positions":86016,"seconds":327.9093069739174,"update":7},{"gradient_norm":0.7470859289169312,"loss":-0.11541590094566345,"mean_reward":null,"positions":98304,"seconds":358.7122434838675,"update":8},{"gradient_norm":0.5518930554389954,"loss":-0.1511152982711792,"mean_reward":null,"positions":110592,"seconds":414.42394595593214,"update":9},{"gradient_norm":1.138114333152771,"loss":-0.19325177371501923,"mean_reward":null,"positions":122880,"seconds":438.2732198978774,"update":10},{"gradient_norm":0.5827913284301758,"loss":-0.15330123901367188,"mean_reward":null,"positions":135168,"seconds":469.7376685049385,"update":11},{"gradient_norm":1.4300124645233154,"loss":-0.9494118690490723,"mean_reward":null,"positions":147456,"seconds":509.5278201010078,"update":12},{"gradient_norm":0.3960418999195099,"loss":-0.11506502330303192,"mean_reward":null,"positions":159744,"seconds":557.4471922670491,"update":13},{"gradient_norm":0.5162414908409119,"loss":-0.11797305196523666,"mean_reward":null,"positions":172032,"seconds":596.7609934129287,"update":14},{"gradient_norm":0.27365973591804504,"loss":-0.037662722170352936,"mean_reward":null,"positions":184320,"seconds":651.7027005348355,"update":15},{"gradient_norm":5.520264625549316,"loss":-1.453701376914978,"mean_reward":null,"positions":196608,"seconds":665.5518124450464,"update":16},{"gradient_norm":1.8220542669296265,"loss":-0.4915573000907898,"mean_reward":null,"positions":208896,"seconds":682.2041524518281,"update":17},{"gradient_norm":0.48240843415260315,"loss":-0.043528638780117035,"mean_reward":null,"positions":221184,"seconds":716.1017245939001,"update":18},{"gradient_norm":0.33644524216651917,"loss":-0.08811994642019272,"mean_reward":null,"positions":233472,"seconds":762.870704401983,"update":19},{"gradient_norm":0.6226396560668945,"loss":-0.10553742200136185,"mean_reward":null,"positions":245760,"seconds":787.8095645778812,"update":20},{"gradient_norm":0.9211729764938354,"loss":-0.18875741958618164,"mean_reward":null,"positions":258048,"seconds":817.8352833669633,"update":21},{"gradient_norm":0.2368898093700409,"loss":-0.04418545961380005,"mean_reward":null,"positions":270336,"seconds":877.498143033823,"update":22},{"gradient_norm":0.5026437044143677,"loss":-0.023194635286927223,"mean_reward":null,"positions":282624,"seconds":905.3091543009505,"update":23},{"gradient_norm":0.5391305685043335,"loss":-0.10577251017093658,"mean_reward":null,"positions":294912,"seconds":956.1653071008623,"update":24},{"gradient_norm":0.6293936371803284,"loss":-0.025131206959486008,"mean_reward":null,"positions":307200,"seconds":977.3168405799661,"update":25},{"gradient_norm":0.7189299464225769,"loss":-0.27981218695640564,"mean_reward":null,"positions":319488,"seconds":1025.6396994688548,"update":26},{"gradient_norm":0.4679880440235138,"loss":-0.009321130812168121,"mean_reward":null,"positions":331776,"seconds":1043.34351826692,"update":27},{"gradient_norm":0.45838218927383423,"loss":-0.10299734771251678,"mean_reward":null,"positions":344064,"seconds":1075.7806981760077,"update":28},{"gradient_norm":0.49755656719207764,"loss":-0.12053272128105164,"mean_reward":null,"positions":356352,"seconds":1134.127101539867,"update":29},{"gradient_norm":0.4340573251247406,"loss":-0.08618558943271637,"mean_reward":null,"positions":368640,"seconds":1178.522843247978,"update":30},{"gradient_norm":0.639819324016571,"loss":-0.19450195133686066,"mean_reward":null,"positions":380928,"seconds":1215.2082774029113,"update":31},{"gradient_norm":0.868110179901123,"loss":-0.31054967641830444,"mean_reward":null,"positions":393216,"seconds":1258.4394343639724,"update":32},{"gradient_norm":1.0689454078674316,"loss":-0.13414013385772705,"mean_reward":null,"positions":405504,"seconds":1281.6050608248916,"update":33},{"gradient_norm":0.5474188923835754,"loss":-0.0857921913266182,"mean_reward":null,"positions":417792,"seconds":1340.3976578768343,"update":34},{"gradient_norm":0.5412011742591858,"loss":-0.04860460013151169,"mean_reward":null,"positions":430080,"seconds":1374.5579588648397,"update":35},{"gradient_norm":0.2309490591287613,"loss":-0.00018738256767392159,"mean_reward":null,"positions":442368,"seconds":1433.3610921090003,"update":36},{"gradient_norm":0.3589034378528595,"loss":-0.01336948573589325,"mean_reward":null,"positions":454656,"seconds":1470.9754410069436,"update":37},{"gradient_norm":0.8566750884056091,"loss":-0.29185178875923157,"mean_reward":null,"positions":466944,"seconds":1530.4939267260488,"update":38},{"gradient_norm":0.45281264185905457,"loss":-0.06886965781450272,"mean_reward":null,"positions":479232,"seconds":1589.3253292129375,"update":39},{"gradient_norm":0.8506408333778381,"loss":-0.03728342056274414,"mean_reward":null,"positions":491520,"seconds":1615.435046989005,"update":40},{"gradient_norm":1.2550318241119385,"loss":-0.12137476354837418,"mean_reward":null,"positions":503808,"seconds":1644.753545989981,"update":41},{"gradient_norm":0.29108449816703796,"loss":-0.011929825879633427,"mean_reward":null,"positions":516096,"seconds":1673.312645799946,"update":42}],"receipt":{"adapter_sha256":"5f160ae39652cc6abe890f87beef90b9896f736af933101f331eea8d76b9be98","peak_allocated_bytes":52718982144,"position_limit":524288,"positions":516096,"recipe":{"accumulation":4,"alpha":32,"beta":0.02,"group_size":2,"learning_rate":9.819805060619657e-06,"method":"opd","micro_batch":2,"rank":16,"rollout_tokens":1024,"seed":7107,"teacher_fraction":0.0,"temperature":0.8},"role":"target","teacher_cache_hash":null,"train_seconds":1676.2896842889022,"training_data_hash":"1118fbf8abc66267f6e0d05201f62f9d4a9b13ee6a8a9626aa4d04ec03aa1284","updates":42}}},"provenance":[{"run":"proxy","job":"allie-f07-math-p5-opd-1009-045951","rollouts_sha256":"076f45895c08e168d5b36e8b4dc4644ea097012379944684206b5439d07ac140","receipt":{"adapter_sha256":"1369a5be794086cc62714d17c689e1b6293b5cfdb30426c3eeb4a1844d5906aa","peak_allocated_bytes":23784590336,"position_limit":524288,"positions":516096,"recipe":{"accumulation":4,"alpha":32,"beta":0.02,"group_size":2,"learning_rate":3e-05,"method":"opd","micro_batch":2,"rank":16,"rollout_tokens":1024,"seed":7107,"teacher_fraction":0.0,"temperature":0.8},"role":"proxy","teacher_cache_hash":null,"train_seconds":979.4048202189151,"training_data_hash":"1118fbf8abc66267f6e0d05201f62f9d4a9b13ee6a8a9626aa4d04ec03aa1284","updates":42}},{"run":"naive","job":"allie-f07-math-target-naive-1009-052437","rollouts_sha256":"4bca7d20854f67720177c0554d797fe7acb335e07ba253300ecc09b7ac05e829","receipt":{"adapter_sha256":"c963b45890ecb83fb8c730bd2d2ef29517aef6ff32e0431ac5a097796c8fcc65","peak_allocated_bytes":52718982144,"position_limit":524288,"positions":516096,"recipe":{"accumulation":4,"alpha":32,"beta":0.02,"group_size":2,"learning_rate":3e-05,"method":"opd","micro_batch":2,"rank":16,"rollout_tokens":1024,"seed":7107,"teacher_fraction":0.0,"temperature":0.8},"role":"target","teacher_cache_hash":null,"train_seconds":1684.7486601730343,"training_data_hash":"1118fbf8abc66267f6e0d05201f62f9d4a9b13ee6a8a9626aa4d04ec03aa1284","updates":42}},{"run":"selected","job":"allie-f07-math-target-selected-1009-052437","rollouts_sha256":"99a1b852aa4ae786fad5d0cf053a5407f889cb0789a31bd1178a9947a1808c36","receipt":{"adapter_sha256":"5f160ae39652cc6abe890f87beef90b9896f736af933101f331eea8d76b9be98","peak_allocated_bytes":52718982144,"position_limit":524288,"positions":516096,"recipe":{"accumulation":4,"alpha":32,"beta":0.02,"group_size":2,"learning_rate":9.819805060619657e-06,"method":"opd","micro_batch":2,"rank":16,"rollout_tokens":1024,"seed":7107,"teacher_fraction":0.0,"temperature":0.8},"role":"target","teacher_cache_hash":null,"train_seconds":1676.2896842889022,"training_data_hash":"1118fbf8abc66267f6e0d05201f62f9d4a9b13ee6a8a9626aa4d04ec03aa1284","updates":42}}],"counts":{"training_problems":512,"opd_problems":42,"student_responses":252,"teacher_responses":512,"teacher_correct":275,"teacher_truncated":36,"teacher_sft_eligible":275},"flags":{"teacher-mismatch":237,"teacher-truncated":36,"rollout-truncated":12,"teacher-right-target-wrong":9,"teacher-wrong-target-right":7,"proxy-wrong-target-right":14,"samples-disagree":8,"missing-reference":1},"debugging_positions":5189632,"trace_grading":"Student answer checks are post-hoc, using the frozen training verifier. OPD did not use these as rewards.","score_availability":"Saved teacher log probabilities score each sampled student token, conditioned on its student prefix, at temperature 0.8. Student log probabilities and full-vocabulary distributions were not saved; exact KL or advantage traces are unavailable."}
