Spaces:
Sleeping
Sleeping
| { | |
| "average_CPS": 59.57754466666667, | |
| "config": { | |
| "model_name": "Qwen/Qwen2.5-32B-Instruct", | |
| "num_fewshot": "10", | |
| "batch_size": 1, | |
| "LANG": "PL", | |
| "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": 60.08, | |
| "stderr": 0.0 | |
| }, | |
| { | |
| "prompt": "p2", | |
| "metric": "f1", | |
| "value": 60.040000000000006, | |
| "stderr": 0.0 | |
| }, | |
| { | |
| "prompt": "p3", | |
| "metric": "f1", | |
| "value": 60.08, | |
| "stderr": 0.0 | |
| } | |
| ], | |
| "average_accuracy": 60.06666666666666, | |
| "best_prompt": 60.08, | |
| "prompt_id": "p1", | |
| "CPS": 60.071989333333335, | |
| "is_dummy": false, | |
| "std_accuracy": 0.023094010767580435 | |
| }, | |
| "RE": { | |
| "prompts": [ | |
| { | |
| "prompt": "p1", | |
| "metric": "f1", | |
| "value": 58.58, | |
| "stderr": 0.0 | |
| }, | |
| { | |
| "prompt": "p2", | |
| "metric": "f1", | |
| "value": 58.68, | |
| "stderr": 0.0 | |
| }, | |
| { | |
| "prompt": "p3", | |
| "metric": "f1", | |
| "value": 59.38, | |
| "stderr": 0.0 | |
| } | |
| ], | |
| "average_accuracy": 58.879999999999995, | |
| "best_prompt": 59.38, | |
| "prompt_id": "p3", | |
| "CPS": 59.083099999999995, | |
| "is_dummy": false, | |
| "std_accuracy": 0.43588989435406944 | |
| }, | |
| "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 | |
| } | |
| } | |
| } |