Upload folder using huggingface_hub

README.md

@@ -104,6 +104,17 @@ substitution, and legacy Tamazight-font mojibake where the sub-dot emphatics
 not carry either defect at measurable rates, but the step is applied regardless so
 the output is guaranteed canonical.
 
+## Baseline Benchmarks
+
+Empirical classification baselines evaluated on `test.jsonl` (1,500 test sentences, 500 per class):
+
+| System / Model | Setting | Accuracy | Macro F1 | Negative F1 | Neutral F1 | Positive F1 |
+|---|---|---|---|---|---|---|
+| **Masinissa-31M** | Linear Probe (Frozen Encoder + Single Linear) | 77.53% | 0.7764 | 0.7389 | 0.8298 | 0.7604 |
+| **Masinissa-31M** | Full Fine-Tuning (RoBERTa-style Head) | **88.80%** | **0.8880** | **0.8831** | **0.9111** | **0.8697** |
+
+Linear Probe evaluates pure feature separability of the frozen pre-trained encoder (Mean Pooling $\rightarrow$ single `nn.Linear`). Full Fine-Tuning fits the encoder end-to-end with a `RobertaClassificationHead` (`Dense` $\rightarrow$ `Tanh` $\rightarrow$ `Dropout` $\rightarrow$ `Linear`).
+
 ## Why this benchmark exists
 
 **No Kabyle sentiment benchmark with a neutral class existed before this release.**