Add Qwen3.5-4B rulewise v9 recall-tune README
Browse files
qwen35_4b_rulewise_v9_recall_tune1/README.md
ADDED
|
@@ -0,0 +1,21 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# Qwen3.5-4B Rulewise V9 Recall Tune Backup
|
| 2 |
+
|
| 3 |
+
This folder archives the recall-oriented continuation branch built from the main Qwen3.5-4B rulewise SFT line on the original ConstructionSite benchmark data.
|
| 4 |
+
|
| 5 |
+
## Contents
|
| 6 |
+
|
| 7 |
+
- `checkpoint-3500`: best checkpoint from the recall-tune continuation branch
|
| 8 |
+
- `final_adapter`: final adapter export from the recall-tune run
|
| 9 |
+
|
| 10 |
+
## Notes
|
| 11 |
+
|
| 12 |
+
- Base checkpoint: `checkpoint-19000` from the main 4B line
|
| 13 |
+
- Training objective: improve R2/R3/R4 recall and reduce misses
|
| 14 |
+
- Continuation setting: lr 3e-5, warmup 0.03, max_steps 3500
|
| 15 |
+
- Rule weights: {1:1.5, 2:6, 3:8, 4:10}
|
| 16 |
+
- Best continuation result: macro F1 0.501, precision 0.766, recall 0.674
|
| 17 |
+
- This branch improves recall but does not surpass the main-line best `checkpoint-19000` in overall macro F1
|
| 18 |
+
|
| 19 |
+
## File policy
|
| 20 |
+
|
| 21 |
+
Each checkpoint keeps the lightweight adapter/trainer artifacts as produced by training. Base model weights are not included.
|