| {"adapter_dir": "checkpoints/step-50/primary", "adapter_sha256": "f0a1e400ab37136131996109ff3c041c8a8e4425e11250c5d82eee098b43baff", "answer_token_accuracy": 0.9217171717171717, "decoded": [{"episode_id": "60134c6e5af0ad59742d", "exact": true, "text": "{\"law\":{\"different\":\"repel\",\"exponent\":1,\"same\":\"attract\"},\"partition\":[0,1,0,1,0,0]}"}, {"episode_id": "d614ab65659bd62b339e", "exact": false, "text": "{\"law\":{\"different\":\"repel\",\"exponent\":1,\"same\":\"attract\"},\"partition\":[0,0,0,1,1,1]}"}, {"episode_id": "1781a89d8e6cd33650a3", "exact": true, "text": "{\"law\":{\"different\":\"repel\",\"exponent\":1,\"same\":\"attract\"},\"partition\":[0,1,1,0,0,1]}"}, {"episode_id": "ccea8fb5cc95448fdfac", "exact": false, "text": "{\"law\":{\"different\":\"repel\",\"exponent\":2,\"same\":\"attract\"},\"partition\":[0,1,0,1,1,0]}"}, {"episode_id": "1ef1104d156c78f54690", "exact": false, "text": "{\"law\":{\"different\":\"repel\",\"exponent\":1,\"same\":\"attract\"},\"partition\":[0,0,1,0,1,0]}"}, {"episode_id": "94ca013f0357f8526974", "exact": false, "text": "{\"law\":{\"different\":\"repel\",\"exponent\":1,\"same\":\"attract\"},\"partition\":[0,1,0,0,1,1]}"}, {"episode_id": "6ecd6b981be1a15eaa46", "exact": true, "text": "{\"law\":{\"different\":\"repel\",\"exponent\":1,\"same\":\"attract\"},\"partition\":[0,1,1,0,0,1]}"}, {"episode_id": "655b63a4c832454e56f6", "exact": false, "text": "{\"law\":{\"different\":\"repel\",\"exponent\":1,\"same\":\"attract\"},\"partition\":[0,0,1,0,1,1]}"}, {"episode_id": "d3cdbeb5336553a8f592", "exact": true, "text": "{\"law\":{\"different\":\"repel\",\"exponent\":1,\"same\":\"attract\"},\"partition\":[0,1,1,1,0,0]}"}, {"episode_id": "92fe86aaca92265e076f", "exact": false, "text": "{\"law\":{\"different\":\"repel\",\"exponent\":1,\"same\":\"attract\"},\"partition\":[0,0,0,1,1,0]}"}, {"episode_id": "6d516a089a32fbbbea1b", "exact": false, "text": "{\"law\":{\"different\":\"repel\",\"exponent\":2,\"same\":\"attract\"},\"partition\":[0,1,0,0,0,1]}"}, {"episode_id": "4573813980de6888fc9d", "exact": false, "text": "{\"law\":{\"different\":\"repel\",\"exponent\":2,\"same\":\"attract\"},\"partition\":[0,1,1,0,1,0]}"}], "examples": 12, "generated_exact_json_accuracy": 0.3333333333333333, "generated_json_parse_rate": 1.0, "step": 50, "teacher_forced_loss": 0.31487923860549927} |
| {"adapter_dir": "checkpoints/step-100/primary", "adapter_sha256": "1d8d7a813be64eae23f083648b24ed18df1813fcca09342641c1fd9549f1749a", "answer_token_accuracy": 0.9621212121212122, "decoded": [{"episode_id": "60134c6e5af0ad59742d", "exact": false, "text": "{\"law\":{\"different\":\"repel\",\"exponent\":2,\"same\":\"repel\"},\"partition\":[0,1,0,1,0,0]}"}, {"episode_id": "d614ab65659bd62b339e", "exact": false, "text": "{\"law\":{\"different\":\"attract\",\"exponent\":1,\"same\":\"attract\"},\"partition\":[0,0,0,1,1,1]}"}, {"episode_id": "1781a89d8e6cd33650a3", "exact": false, "text": "{\"law\":{\"different\":\"repel\",\"exponent\":2,\"same\":\"repel\"},\"partition\":[0,1,1,0,0,1]}"}, {"episode_id": "ccea8fb5cc95448fdfac", "exact": false, "text": "{\"law\":{\"different\":\"attract\",\"exponent\":1,\"same\":\"attract\"},\"partition\":[0,1,0,1,1,0]}"}, {"episode_id": "1ef1104d156c78f54690", "exact": false, "text": "{\"law\":{\"different\":\"repel\",\"exponent\":1,\"same\":\"attract\"},\"partition\":[0,0,1,0,1,1]}"}, {"episode_id": "94ca013f0357f8526974", "exact": true, "text": "{\"law\":{\"different\":\"repel\",\"exponent\":2,\"same\":\"repel\"},\"partition\":[0,1,0,0,1,1]}"}, {"episode_id": "6ecd6b981be1a15eaa46", "exact": false, "text": "{\"law\":{\"different\":\"repel\",\"exponent\":2,\"same\":\"repel\"},\"partition\":[0,1,1,0,0,1]}"}, {"episode_id": "655b63a4c832454e56f6", "exact": false, "text": "{\"law\":{\"different\":\"repel\",\"exponent\":2,\"same\":\"attract\"},\"partition\":[0,0,1,0,1,1]}"}, {"episode_id": "d3cdbeb5336553a8f592", "exact": true, "text": "{\"law\":{\"different\":\"repel\",\"exponent\":1,\"same\":\"attract\"},\"partition\":[0,1,1,1,0,0]}"}, {"episode_id": "92fe86aaca92265e076f", "exact": true, "text": "{\"law\":{\"different\":\"repel\",\"exponent\":2,\"same\":\"repel\"},\"partition\":[0,0,0,1,1,0]}"}, {"episode_id": "6d516a089a32fbbbea1b", "exact": false, "text": "{\"law\":{\"different\":\"repel\",\"exponent\":2,\"same\":\"attract\"},\"partition\":[0,1,0,0,0,1]}"}, {"episode_id": "4573813980de6888fc9d", "exact": false, "text": "{\"law\":{\"different\":\"repel\",\"exponent\":2,\"same\":\"attract\"},\"partition\":[0,1,1,0,1,0]}"}], "examples": 12, "generated_exact_json_accuracy": 0.25, "generated_json_parse_rate": 1.0, "step": 100, "teacher_forced_loss": 0.05907239951193333} |
| {"adapter_dir": "checkpoints/step-200/primary", "adapter_sha256": "694e7f0120625822365ea58e6cb2c1808579fd4a4e0977b676199ba996c34588", "answer_token_accuracy": 0.9671717171717171, "decoded": [{"episode_id": "60134c6e5af0ad59742d", "exact": false, "text": "{\"law\":{\"different\":\"attract\",\"exponent\":1,\"same\":\"attract\"},\"partition\":[0,1,0,1,0,0]}"}, {"episode_id": "d614ab65659bd62b339e", "exact": false, "text": "{\"law\":{\"different\":\"attract\",\"exponent\":2,\"same\":\"attract\"},\"partition\":[0,0,0,1,1,1]}"}, {"episode_id": "1781a89d8e6cd33650a3", "exact": false, "text": "{\"law\":{\"different\":\"repel\",\"exponent\":2,\"same\":\"repel\"},\"partition\":[0,1,1,0,0,1]}"}, {"episode_id": "ccea8fb5cc95448fdfac", "exact": false, "text": "{\"law\":{\"different\":\"attract\",\"exponent\":1,\"same\":\"attract\"},\"partition\":[0,1,0,1,1,0]}"}, {"episode_id": "1ef1104d156c78f54690", "exact": true, "text": "{\"law\":{\"different\":\"attract\",\"exponent\":1,\"same\":\"attract\"},\"partition\":[0,0,1,0,1,1]}"}, {"episode_id": "94ca013f0357f8526974", "exact": false, "text": "{\"law\":{\"different\":\"attract\",\"exponent\":1,\"same\":\"attract\"},\"partition\":[0,1,0,0,1,1]}"}, {"episode_id": "6ecd6b981be1a15eaa46", "exact": false, "text": "{\"law\":{\"different\":\"repel\",\"exponent\":2,\"same\":\"repel\"},\"partition\":[0,1,1,0,0,1]}"}, {"episode_id": "655b63a4c832454e56f6", "exact": false, "text": "{\"law\":{\"different\":\"repel\",\"exponent\":1,\"same\":\"attract\"},\"partition\":[0,0,1,0,1,1]}"}, {"episode_id": "d3cdbeb5336553a8f592", "exact": false, "text": "{\"law\":{\"different\":\"attract\",\"exponent\":2,\"same\":\"repel\"},\"partition\":[0,1,1,1,0,0]}"}, {"episode_id": "92fe86aaca92265e076f", "exact": false, "text": "{\"law\":{\"different\":\"repel\",\"exponent\":1,\"same\":\"attract\"},\"partition\":[0,0,0,1,1,0]}"}, {"episode_id": "6d516a089a32fbbbea1b", "exact": false, "text": "{\"law\":{\"different\":\"attract\",\"exponent\":2,\"same\":\"repel\"},\"partition\":[0,1,0,0,0,1]}"}, {"episode_id": "4573813980de6888fc9d", "exact": false, "text": "{\"law\":{\"different\":\"repel\",\"exponent\":1,\"same\":\"attract\"},\"partition\":[0,1,1,0,1,0]}"}], "examples": 12, "generated_exact_json_accuracy": 0.08333333333333333, "generated_json_parse_rate": 1.0, "step": 200, "teacher_forced_loss": 0.06225178204476833} |
|
|
