Update README.md

README.md

@@ -95,14 +95,31 @@ Level 0: 762 (27.2%) • Level 1: 674 (24.1%) • Level 2: 795 (28.4%) • Level
 - Class balancing: sqrt‑scaled class weights + weighted random sampler
 
 ## Evaluation Results
-Test Set (400 held‑out samples)
-
-| Metric                   | Value |
-|--------------------------|-------|
-| Exact Match Accuracy     | 65.0% |
-| Adjacent (±1) Accuracy   | 92.5% |
-| Macro F1                 | 0.663 |
-| Weighted F1              | 0.653 |
+
+Reported on 600 held‑out samples from TRAINING.jsonl (internal test).
+|Metric|Value|
+|----|----|
+|Exact Match Accuracy|64.5%|
+|Adjacent (±1) Accuracy|93.0%|
+|Macro F1|0.663|
+|Weighted F1|0.653|
+
+Per‑Class Performance (internal test, 600 samples)
+|Level|Precision|Recall|F1|Support|
+|----|----|----|----|----|
+|L0 (Trivial)|0.658|0.626|0.642|163|
+|L1 (Simple)|0.457|0.628|0.529|145|
+|L2 (Moderate)|0.683|0.571|0.622|170|
+|L3 (Complex)|0.933|0.795|0.858|122|
+
+Confusion Matrix (internal test, 600 samples)
+
+|    |Pred L0|Pred L1|Pred L2|Pred L3|
+|----|----|----|----|----|
+|True L0|102|46|13|2|
+|True L1|35|91|18|1|
+|True L2|15|54|97|4|
+|True L3|3|8|14|97|
 
 ## How to Get Started with the Model