| { |
| "evaluated_at": "2026-07-02T13:15:23+00:00", |
| "base_model": "mlx-community/Qwen2.5-0.5B-Instruct-4bit", |
| "task": "json-extractor-demo", |
| "adapter": null, |
| "test_set": "runs/json-extractor-demo/data/test.jsonl", |
| "decoding": { |
| "temperature": 0.0, |
| "max_tokens": 120, |
| "chat_template_kwargs": {} |
| }, |
| "metrics": { |
| "examples": 40, |
| "exact_match": 0.0, |
| "json_examples": 40, |
| "json_validity": 1.0, |
| "field_accuracy": 0.1667, |
| "mean_latency_s": 0.133 |
| }, |
| "samples": [ |
| { |
| "expected": "{\"title\": \"Buy a birthday gift\", \"tags\": [\"shopping\", \"tomorrow\"], \"priority\": \"normal\"}", |
| "generated": "{\n \"title\": \"buy a birthday gift tomorrow\",\n \"tags\": [\"todo\"],\n \"priority\": \"high\"\n}" |
| }, |
| { |
| "expected": "{\"title\": \"Buy a birthday gift\", \"tags\": [\"shopping\", \"next-week\"], \"priority\": \"normal\"}", |
| "generated": "{\n \"title\": \"buy a birthday gift next week\",\n \"tags\": [\"birthday\", \"gift\", \"next week\"],\n \"priority\": \"high\"\n}" |
| }, |
| { |
| "expected": "{\"title\": \"Fix the kitchen tap\", \"tags\": [\"home\", \"friday\"], \"priority\": \"normal\"}", |
| "generated": "{\n \"title\": \"fix the kitchen tap before friday\",\n \"tags\": [\"kitchen\", \"tap\", \"fix\"],\n \"priority\": \"high\"\n}" |
| }, |
| { |
| "expected": "{\"title\": \"Pay the electricity bill\", \"tags\": [\"finance\", \"today\"], \"priority\": \"high\"}", |
| "generated": "{\n \"title\": \"pay the electricity bill today\",\n \"tags\": [\"high\"],\n \"priority\": \"high\"\n}" |
| }, |
| { |
| "expected": "{\"title\": \"Clean the garage\", \"tags\": [\"home\", \"tomorrow\"], \"priority\": \"normal\"}", |
| "generated": "{\n \"title\": \"need to clean the garage tomorrow\",\n \"tags\": [\"garage\", \"cleaning\"],\n \"priority\": \"high\"\n}" |
| } |
| ] |
| } |
|
|
