| { |
| "schema_version": "0.2.2", |
| "evaluation_id": "math_500|0/CohereLabs/command-a-plus-05-2026-bf16/1783690408.011777", |
| "evaluation_timestamp": "777717", |
| "retrieved_timestamp": "1783690408.011777", |
| "source_metadata": { |
| "source_name": "lighteval", |
| "source_type": "evaluation_run", |
| "source_organization_name": "RedHatAI", |
| "evaluator_relationship": "third_party" |
| }, |
| "eval_library": { |
| "name": "lighteval", |
| "version": "0.13.1.dev0" |
| }, |
| "model_info": { |
| "name": "CohereLabs/command-a-plus-05-2026-bf16", |
| "id": "CohereLabs/command-a-plus-05-2026-bf16", |
| "developer": "CohereLabs", |
| "inference_engine": { |
| "name": "vllm", |
| "version": "0.22.1" |
| }, |
| "additional_details": { |
| "provider": "hosted_vllm", |
| "base_url": "http://localhost:8000/v1", |
| "concurrent_requests": "16", |
| "verbose": "False", |
| "max_model_length": "131072", |
| "api_max_retry": "8", |
| "api_retry_sleep": "1.0", |
| "api_retry_multiplier": "2.0", |
| "timeout": "3600.0", |
| "num_seeds_merged": "3" |
| } |
| }, |
| "evaluation_results": [ |
| { |
| "evaluation_name": "math_500", |
| "source_data": { |
| "dataset_name": "math_500", |
| "source_type": "hf_dataset", |
| "hf_repo": "HuggingFaceH4/MATH-500", |
| "hf_split": "test" |
| }, |
| "metric_config": { |
| "evaluation_description": "pass@k:k=1&n=1", |
| "lower_is_better": false, |
| "score_type": "continuous", |
| "min_score": 0.0, |
| "max_score": 1.0 |
| }, |
| "score_details": { |
| "score": 0.898, |
| "details": { |
| "seed_scores": "[0.902, 0.882, 0.91]", |
| "evaluation_timestamps": "[782779, 787744, 777717]", |
| "seed_values": "[1, 2, 0]" |
| }, |
| "uncertainty": { |
| "standard_error": { |
| "value": 0.008326663997864539, |
| "method": "across_seeds" |
| }, |
| "num_samples": 3 |
| } |
| }, |
| "generation_config": { |
| "generation_args": { |
| "temperature": 0.6, |
| "top_p": 0.95, |
| "max_tokens": 65536, |
| "max_attempts": 1 |
| }, |
| "additional_details": { |
| "seed": "1", |
| "num_fewshot": "0" |
| } |
| } |
| } |
| ] |
| } |
