Spaces:
Running
Running
Upload folder using huggingface_hub
Browse files
artifacts/runs/20260426-033957-corrected-eval/corrected_metrics.json
ADDED
|
@@ -0,0 +1,30 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"task_ids": [
|
| 3 |
+
"easy_syntax_fix",
|
| 4 |
+
"medium_logic_fix",
|
| 5 |
+
"hard_multi_bug",
|
| 6 |
+
"hard_finance_explosion"
|
| 7 |
+
],
|
| 8 |
+
"samples_per_task": 8,
|
| 9 |
+
"per_task_baseline_reward": {
|
| 10 |
+
"easy_syntax_fix": 0.0,
|
| 11 |
+
"medium_logic_fix": 0.0,
|
| 12 |
+
"hard_multi_bug": 0.0,
|
| 13 |
+
"hard_finance_explosion": 0.0
|
| 14 |
+
},
|
| 15 |
+
"per_task_post_reward": {
|
| 16 |
+
"easy_syntax_fix": 0.0,
|
| 17 |
+
"medium_logic_fix": 0.0,
|
| 18 |
+
"hard_multi_bug": 0.0,
|
| 19 |
+
"hard_finance_explosion": 0.0
|
| 20 |
+
},
|
| 21 |
+
"delta_per_task": {
|
| 22 |
+
"easy_syntax_fix": 0.0,
|
| 23 |
+
"medium_logic_fix": 0.0,
|
| 24 |
+
"hard_multi_bug": 0.0,
|
| 25 |
+
"hard_finance_explosion": 0.0
|
| 26 |
+
},
|
| 27 |
+
"baseline_avg_reward": 0.0,
|
| 28 |
+
"post_avg_reward": 0.0,
|
| 29 |
+
"delta_avg_reward": 0.0
|
| 30 |
+
}
|
artifacts/runs/20260426-033957-corrected-eval/performance_comparison_corrected.png
ADDED
|
artifacts/runs/20260426-033957-corrected-eval/task_delta_corrected.png
ADDED
|