Upload eval_tuned.json with huggingface_hub
Browse files- eval_tuned.json +42 -0
eval_tuned.json
ADDED
|
@@ -0,0 +1,42 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"evaluated_at": "2026-07-02T13:15:31+00:00",
|
| 3 |
+
"base_model": "mlx-community/Qwen2.5-0.5B-Instruct-4bit",
|
| 4 |
+
"task": "json-extractor-demo",
|
| 5 |
+
"adapter": "runs/json-extractor-demo/adapters",
|
| 6 |
+
"test_set": "runs/json-extractor-demo/data/test.jsonl",
|
| 7 |
+
"decoding": {
|
| 8 |
+
"temperature": 0.0,
|
| 9 |
+
"max_tokens": 120,
|
| 10 |
+
"chat_template_kwargs": {}
|
| 11 |
+
},
|
| 12 |
+
"metrics": {
|
| 13 |
+
"examples": 40,
|
| 14 |
+
"exact_match": 1.0,
|
| 15 |
+
"json_examples": 40,
|
| 16 |
+
"json_validity": 1.0,
|
| 17 |
+
"field_accuracy": 1.0,
|
| 18 |
+
"mean_latency_s": 0.164
|
| 19 |
+
},
|
| 20 |
+
"samples": [
|
| 21 |
+
{
|
| 22 |
+
"expected": "{\"title\": \"Buy a birthday gift\", \"tags\": [\"shopping\", \"tomorrow\"], \"priority\": \"normal\"}",
|
| 23 |
+
"generated": "{\"title\": \"Buy a birthday gift\", \"tags\": [\"shopping\", \"tomorrow\"], \"priority\": \"normal\"}"
|
| 24 |
+
},
|
| 25 |
+
{
|
| 26 |
+
"expected": "{\"title\": \"Buy a birthday gift\", \"tags\": [\"shopping\", \"next-week\"], \"priority\": \"normal\"}",
|
| 27 |
+
"generated": "{\"title\": \"Buy a birthday gift\", \"tags\": [\"shopping\", \"next-week\"], \"priority\": \"normal\"}"
|
| 28 |
+
},
|
| 29 |
+
{
|
| 30 |
+
"expected": "{\"title\": \"Fix the kitchen tap\", \"tags\": [\"home\", \"friday\"], \"priority\": \"normal\"}",
|
| 31 |
+
"generated": "{\"title\": \"Fix the kitchen tap\", \"tags\": [\"home\", \"friday\"], \"priority\": \"normal\"}"
|
| 32 |
+
},
|
| 33 |
+
{
|
| 34 |
+
"expected": "{\"title\": \"Pay the electricity bill\", \"tags\": [\"finance\", \"today\"], \"priority\": \"high\"}",
|
| 35 |
+
"generated": "{\"title\": \"Pay the electricity bill\", \"tags\": [\"finance\", \"today\"], \"priority\": \"high\"}"
|
| 36 |
+
},
|
| 37 |
+
{
|
| 38 |
+
"expected": "{\"title\": \"Clean the garage\", \"tags\": [\"home\", \"tomorrow\"], \"priority\": \"normal\"}",
|
| 39 |
+
"generated": "{\"title\": \"Clean the garage\", \"tags\": [\"home\", \"tomorrow\"], \"priority\": \"normal\"}"
|
| 40 |
+
}
|
| 41 |
+
]
|
| 42 |
+
}
|