| {"task_name": "hellaswag", "task_hash": "8312d0c6fac4c6da5cc98a431402ea60", "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": "hellaswag", "task_core": "hellaswag", "limit": 1000, "split": "validation", "num_shots": 5, "fewshot_seed": 1234, "primary_metric": "acc_per_char", "random_subsample_seed": 1234, "context_kwargs": null, "generation_kwargs": null, "metric_kwargs": {"uncond_docid_offset": 1000000}, "native_id_field": "ind", "fewshot_source": "OLMES:hellaswag", "dataset_path": "hellaswag", "dataset_name": null, "use_chat_format": null, "version": 0, "revision": null, "metadata": {"regimes": ["OLMES-v0.1"], "alias": "hellaswag: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": 125.9155502319336, "current_date": "2024-11-19 21:17:28 UTC", "num_instances": 1000, "beaker_info": {}, "metrics": {"acc_raw": 0.469, "acc_per_token": 0.592, "acc_per_char": 0.59, "correct_loss_raw": 72.27628855609893, "incorrect_loss_raw": 89.32188357480365, "correct_loss_per_token": 2.3943869034698344, "incorrect_loss_per_token": 2.9809331095903073, "correct_loss_per_char": 0.5284280503451793, "incorrect_loss_per_char": 0.6609210587038847, "acc_uncond": 0.478, "correct_loss_uncond": -26.180681183815004, "incorrect_loss_uncond": -20.375546882629408, "primary_score": 0.59}, "task_idx": 9} |