File size: 3,609 Bytes
2616764 | 1 2 | {"gold_dir": "results/composition/op-difficulty-10B/graphs", "generated_path": "results/Qwen3-100M-midtrain/composition", "results_dir": "results/Qwen3-100M-midtrain/composition/graphs", "per_file": [{"file": "results/Qwen3-100M-midtrain/composition/checkpoint-38797_id_generations.jsonl", "total": 1152000, "missing_solution": 0, "missing_gold": 0, "parse_fail": 0, "perfect": 1024558, "value_mismatch": 8267, "dependency_mismatch": 126000, "answer_mismatch": 7374, "extra_nodes": 11272, "missing_nodes": 8614, "perfect_with_extra": 1191, "unused_gold_entries": 10000, "template_stats": {"crazy_zootopia": {"total": 379008, "perfect": 337758, "value_mismatch": 2544, "dependency_mismatch": 40803, "answer_mismatch": 2204, "extra_nodes": 3484, "missing_nodes": 2675, "perfect_with_extra": 419}, "teachers_in_school": {"total": 390656, "perfect": 347066, "extra_nodes": 3898, "perfect_with_extra": 379, "dependency_mismatch": 43092, "value_mismatch": 3113, "answer_mismatch": 2966, "missing_nodes": 3187}, "movie_festival_awards": {"total": 382336, "perfect": 339734, "extra_nodes": 3890, "perfect_with_extra": 393, "value_mismatch": 2610, "dependency_mismatch": 42105, "answer_mismatch": 2204, "missing_nodes": 2752}}, "results_path": "results/Qwen3-100M-midtrain/composition/graphs/checkpoint-38797_id_generations.jsonl.graph_results.jsonl"}, {"file": "results/Qwen3-100M-midtrain/composition/checkpoint-38797_ood_generations.jsonl", "total": 1280000, "missing_solution": 0, "missing_gold": 0, "parse_fail": 0, "perfect": 656373, "value_mismatch": 203587, "dependency_mismatch": 603266, "answer_mismatch": 154019, "extra_nodes": 43456, "missing_nodes": 133737, "perfect_with_extra": 209, "unused_gold_entries": 9000, "template_stats": {"crazy_zootopia": {"total": 430720, "perfect": 220371, "dependency_mismatch": 203256, "value_mismatch": 69652, "answer_mismatch": 53878, "extra_nodes": 16415, "perfect_with_extra": 56, "missing_nodes": 48414}, "movie_festival_awards": {"total": 418816, "perfect": 214999, "extra_nodes": 13360, "perfect_with_extra": 81, "value_mismatch": 69881, "dependency_mismatch": 197311, "answer_mismatch": 52549, "missing_nodes": 43418}, "teachers_in_school": {"total": 430464, "perfect": 221003, "dependency_mismatch": 202699, "extra_nodes": 13681, "perfect_with_extra": 72, "value_mismatch": 64054, "answer_mismatch": 47592, "missing_nodes": 41905}}, "results_path": "results/Qwen3-100M-midtrain/composition/graphs/checkpoint-38797_ood_generations.jsonl.graph_results.jsonl"}], "aggregate": {"files": 2, "total": 2432000, "missing_solution": 0, "missing_gold": 0, "parse_fail": 0, "perfect": 1680931, "value_mismatch": 211854, "dependency_mismatch": 729266, "answer_mismatch": 161393, "extra_nodes": 54728, "missing_nodes": 142351, "perfect_with_extra": 1400, "skipped": 0, "unused_gold_entries": 0, "aggregated": true, "report_path": "results/Qwen3-100M-midtrain/composition/graphs/composition_score_report.jsonl", "template_stats": {"crazy_zootopia": {"total": 809728, "perfect": 558129, "value_mismatch": 72196, "dependency_mismatch": 244059, "answer_mismatch": 56082, "extra_nodes": 19899, "missing_nodes": 51089, "perfect_with_extra": 475}, "teachers_in_school": {"total": 821120, "perfect": 568069, "extra_nodes": 17579, "perfect_with_extra": 451, "dependency_mismatch": 245791, "value_mismatch": 67167, "answer_mismatch": 50558, "missing_nodes": 45092}, "movie_festival_awards": {"total": 801152, "perfect": 554733, "extra_nodes": 17250, "perfect_with_extra": 474, "value_mismatch": 72491, "dependency_mismatch": 239416, "answer_mismatch": 54753, "missing_nodes": 46170}}}}
|