{"task_name": "openbookqa", "task_hash": "bcd3c6e0e23954870d75bd4cd800afc9", "model_hash": "03418cf8091a9882619950ffb07429a5", "model_config": {"model": "hf_checkpoints/dclm-pool-1b-1x-h-fasttext-open_lm_1b_swiglutorch-warm5000-lr0p003-wd0p033-cd3e-05-bs256-mult1-seed124-tokens28795904000", "revision": null, "trust_remote_code": null, "max_length": 2048, "model_path": null, "model_type": "hf"}, "task_config": {"task_name": "openbookqa", "task_core": "openbookqa", "limit": null, "split": "test", "num_shots": 5, "fewshot_seed": 1234, "primary_metric": "acc_uncond", "random_subsample_seed": 1234, "context_kwargs": {"no_prefix": false}, "generation_kwargs": null, "metric_kwargs": {"uncond_docid_offset": 1000000}, "native_id_field": "id", "fewshot_source": "OLMES:openbookqa", "dataset_path": "openbookqa", "dataset_name": "main", "use_chat_format": null, "version": 0, "revision": null, "metadata": {"regimes": ["OLMES-v0.1"], "alias": "openbookqa:rc::olmes"}}, "compute_config": {"batch_size": "16", "max_batch_size": 32, "output_dir": "hf_checkpoints/dclm-pool-1b-1x-h-fasttext-open_lm_1b_swiglutorch-warm5000-lr0p003-wd0p033-cd3e-05-bs256-mult1-seed124-tokens28795904000/olmes_fp32/core_9mcqa", "num_recorded_inputs": 3, "save_raw_requests": true, "recompute_metrics": false, "check_datalake": false, "push_datalake": false, "no_datalake": false}, "processing_time": 13.533832550048828, "current_date": "2024-11-19 21:19:43 UTC", "num_instances": 500, "beaker_info": {}, "metrics": {"acc_raw": 0.258, "acc_per_token": 0.382, "acc_per_char": 0.344, "correct_loss_raw": 15.555774765849113, "incorrect_loss_raw": 14.663584057887403, "correct_loss_per_token": 4.5521495499962, "incorrect_loss_per_token": 5.301223985967573, "correct_loss_per_char": 0.8821060165202664, "incorrect_loss_per_char": 0.9975226397142319, "acc_uncond": 0.488, "correct_loss_uncond": -9.155967952847481, "incorrect_loss_uncond": -7.157924043258035, "primary_score": 0.488}, "task_idx": 11} |