Spaces:
Sleeping
Sleeping
| { | |
| "average_CPS": 15.684724516666668, | |
| "config": { | |
| "model_name": "Qwen/Qwen2.5-32B-Instruct", | |
| "num_fewshot": "0", | |
| "batch_size": 1, | |
| "LANG": "IT", | |
| "model": "Qwen/Qwen2.5-32B-Instruct", | |
| "base_model": "Qwen2ForCausalLM", | |
| "revision": "5ede1c97bbab6ce5cda5812749b4c0bdf79b18dd", | |
| "submitted_time": "2024-09-17 04:17:55+00:00", | |
| "num_params_billion": 32.763876352, | |
| "language": "en" | |
| }, | |
| "tasks": { | |
| "NER": { | |
| "prompts": [ | |
| { | |
| "prompt": "p1", | |
| "metric": "f1", | |
| "value": 37.580000000000005, | |
| "stderr": 0.0 | |
| }, | |
| { | |
| "prompt": "p2", | |
| "metric": "f1", | |
| "value": 16.470000000000002, | |
| "stderr": 0.0 | |
| }, | |
| { | |
| "prompt": "p3", | |
| "metric": "f1", | |
| "value": 27.96, | |
| "stderr": 0.0 | |
| } | |
| ], | |
| "average_accuracy": 27.336666666666673, | |
| "best_prompt": 37.580000000000005, | |
| "prompt_id": "p1", | |
| "CPS": 33.730555333333335, | |
| "is_dummy": false, | |
| "std_accuracy": 10.568795264046578 | |
| }, | |
| "RE": { | |
| "prompts": [ | |
| { | |
| "prompt": "p1", | |
| "metric": "f1", | |
| "value": 45.050000000000004, | |
| "stderr": 0.0 | |
| }, | |
| { | |
| "prompt": "p2", | |
| "metric": "f1", | |
| "value": 41.589999999999996, | |
| "stderr": 0.0 | |
| }, | |
| { | |
| "prompt": "p3", | |
| "metric": "f1", | |
| "value": 44.47, | |
| "stderr": 0.0 | |
| } | |
| ], | |
| "average_accuracy": 43.70333333333334, | |
| "best_prompt": 45.050000000000004, | |
| "prompt_id": "p1", | |
| "CPS": 44.44332666666667, | |
| "is_dummy": false, | |
| "std_accuracy": 1.8530335488957954 | |
| }, | |
| "RML": { | |
| "prompts": [ | |
| { | |
| "prompt": "p1", | |
| "metric": "f1", | |
| "value": 0.0, | |
| "stderr": 0.0 | |
| }, | |
| { | |
| "prompt": "p1", | |
| "metric": "f1", | |
| "value": 0.0, | |
| "stderr": 0.0 | |
| }, | |
| { | |
| "prompt": "p2", | |
| "metric": "f1", | |
| "value": 0.16999999999999998, | |
| "stderr": 0.0 | |
| }, | |
| { | |
| "prompt": "p3", | |
| "metric": "f1", | |
| "value": 0.0, | |
| "stderr": 0.0 | |
| } | |
| ], | |
| "average_accuracy": 0.042499999999999996, | |
| "best_prompt": 0.16999999999999998, | |
| "prompt_id": "p2", | |
| "CPS": 0.16978324999999997, | |
| "is_dummy": false, | |
| "std_accuracy": 0.08499999999999999 | |
| }, | |
| "DIA": { | |
| "prompts": [ | |
| { | |
| "prompt": "p1", | |
| "metric": "f1", | |
| "value": 0.0, | |
| "stderr": 0.0 | |
| }, | |
| { | |
| "prompt": "p2", | |
| "metric": "f1", | |
| "value": 0.0, | |
| "stderr": 0.0 | |
| }, | |
| { | |
| "prompt": "p3", | |
| "metric": "f1", | |
| "value": 0.0, | |
| "stderr": 0.0 | |
| } | |
| ], | |
| "average_accuracy": 0.0, | |
| "best_prompt": 0.0, | |
| "prompt_id": "p1", | |
| "CPS": 0.0, | |
| "is_dummy": false, | |
| "std_accuracy": 0.0 | |
| }, | |
| "HIS": { | |
| "prompts": [ | |
| { | |
| "prompt": "p1", | |
| "metric": "f1", | |
| "value": 0.0, | |
| "stderr": 0.0 | |
| }, | |
| { | |
| "prompt": "p2", | |
| "metric": "f1", | |
| "value": 0.0, | |
| "stderr": 0.0 | |
| }, | |
| { | |
| "prompt": "p3", | |
| "metric": "f1", | |
| "value": 0.08, | |
| "stderr": 0.0 | |
| } | |
| ], | |
| "average_accuracy": 0.02666666666666667, | |
| "best_prompt": 0.08, | |
| "prompt_id": "p3", | |
| "CPS": 0.07995733333333332, | |
| "is_dummy": false, | |
| "std_accuracy": 0.046188021535170064 | |
| } | |
| } | |
| } |