Spaces:
Sleeping
Sleeping
| { | |
| "average_CPS": 27.026387333333332, | |
| "config": { | |
| "model_name": "Qwen/Qwen2.5-14B-Instruct-1M", | |
| "num_fewshot": "0", | |
| "batch_size": 1, | |
| "LANG": "SK", | |
| "model": "Qwen/Qwen2.5-14B-Instruct-1M", | |
| "base_model": "Qwen2ForCausalLM", | |
| "revision": "620fad32de7bdd2293b3d99b39eba2fe63e97438", | |
| "submitted_time": "2025-01-23 13:23:24+00:00", | |
| "num_params_billion": 14.770033664, | |
| "language": "en" | |
| }, | |
| "tasks": { | |
| "NER": { | |
| "prompts": [ | |
| { | |
| "prompt": "p1", | |
| "metric": "f1", | |
| "value": 12.2, | |
| "stderr": 0.0 | |
| }, | |
| { | |
| "prompt": "p2", | |
| "metric": "f1", | |
| "value": 4.26, | |
| "stderr": 0.0 | |
| }, | |
| { | |
| "prompt": "p3", | |
| "metric": "f1", | |
| "value": 12.2, | |
| "stderr": 0.0 | |
| } | |
| ], | |
| "average_accuracy": 9.553333333333333, | |
| "best_prompt": 12.2, | |
| "prompt_id": "p1", | |
| "CPS": 11.877106666666666, | |
| "is_dummy": false, | |
| "std_accuracy": 4.584161137365628 | |
| }, | |
| "RE": { | |
| "prompts": [ | |
| { | |
| "prompt": "p1", | |
| "metric": "f1", | |
| "value": 40.27, | |
| "stderr": 0.0 | |
| }, | |
| { | |
| "prompt": "p2", | |
| "metric": "f1", | |
| "value": 42.94, | |
| "stderr": 0.0 | |
| }, | |
| { | |
| "prompt": "p3", | |
| "metric": "f1", | |
| "value": 40.27, | |
| "stderr": 0.0 | |
| } | |
| ], | |
| "average_accuracy": 41.160000000000004, | |
| "best_prompt": 42.94, | |
| "prompt_id": "p2", | |
| "CPS": 42.175668, | |
| "is_dummy": false, | |
| "std_accuracy": 1.5415252187362976 | |
| }, | |
| "RML": { | |
| "prompts": [ | |
| { | |
| "prompt": "p1", | |
| "metric": "f1", | |
| "value": 0.0, | |
| "stderr": 0.0 | |
| } | |
| ], | |
| "average_accuracy": null, | |
| "std_accuracy": null, | |
| "best_prompt": null, | |
| "prompt_id": null, | |
| "CPS": null, | |
| "is_dummy": true | |
| }, | |
| "HIS": { | |
| "prompts": [ | |
| { | |
| "prompt": "p1", | |
| "metric": "f1", | |
| "value": 0.0, | |
| "stderr": 0.0 | |
| } | |
| ], | |
| "average_accuracy": null, | |
| "std_accuracy": null, | |
| "best_prompt": null, | |
| "prompt_id": null, | |
| "CPS": null, | |
| "is_dummy": true | |
| }, | |
| "DIA": { | |
| "prompts": [ | |
| { | |
| "prompt": "p1", | |
| "metric": "f1", | |
| "value": 0.0, | |
| "stderr": 0.0 | |
| } | |
| ], | |
| "average_accuracy": null, | |
| "std_accuracy": null, | |
| "best_prompt": null, | |
| "prompt_id": null, | |
| "CPS": null, | |
| "is_dummy": true | |
| } | |
| } | |
| } |