trainer-runtime e2e job_e2e_1774221797629 (metadata)
Browse files
artifacts/evaluation/result.json
ADDED
|
@@ -0,0 +1,38 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"model": "/output/job_e2e_1774221797629/merged/trainer-e2e-job_e2e_1774221797629",
|
| 3 |
+
"metrics": {
|
| 4 |
+
"model": "/output/job_e2e_1774221797629/merged/trainer-e2e-job_e2e_1774221797629",
|
| 5 |
+
"samples": 1,
|
| 6 |
+
"parseSuccessRate": 0.0,
|
| 7 |
+
"mae": null,
|
| 8 |
+
"rmse": null,
|
| 9 |
+
"exactRate": 0.0,
|
| 10 |
+
"within1Rate": 0.0,
|
| 11 |
+
"within2Rate": 0.0,
|
| 12 |
+
"meanSignedError": null,
|
| 13 |
+
"avgPredictedScore": null,
|
| 14 |
+
"parseErrors": 1,
|
| 15 |
+
"inferenceErrors": 1,
|
| 16 |
+
"emptyResponses": 1
|
| 17 |
+
},
|
| 18 |
+
"rows": [
|
| 19 |
+
{
|
| 20 |
+
"sampleId": "eval_1",
|
| 21 |
+
"question": "Сколько будет 2+2?",
|
| 22 |
+
"candidateAnswer": "4",
|
| 23 |
+
"referenceScore": 5.0,
|
| 24 |
+
"maxScore": 5,
|
| 25 |
+
"hashTags": [
|
| 26 |
+
"math",
|
| 27 |
+
"sanity"
|
| 28 |
+
],
|
| 29 |
+
"predictedScore": null,
|
| 30 |
+
"predictedFeedback": null,
|
| 31 |
+
"rawResponse": "",
|
| 32 |
+
"parseError": true,
|
| 33 |
+
"inferenceError": true,
|
| 34 |
+
"absoluteError": null,
|
| 35 |
+
"error": "'Qwen2Attention' object has no attribute 'apply_qkv'"
|
| 36 |
+
}
|
| 37 |
+
]
|
| 38 |
+
}
|