Normalise 34 rows that kept homoglyphs; correct the card's class supports

README.md

@@ -106,14 +106,28 @@ the output is guaranteed canonical.
 
 ## Baseline Benchmarks
 
-Empirical classification baselines evaluated on `test.jsonl` (1,500 test sentences, 500 per class):
+Scored on `test.jsonl` — 1,500 sentences, 521 negative / 490 neutral / 489 positive.
+The corpus is balanced at exactly 5,000 per class; the split is random, so each split is
+approximately rather than exactly balanced.
 
 | 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** |
+| **Masinissa-31M** | linear probe, frozen encoder | 77.53% | 0.7764 | 0.7389 | 0.8298 | 0.7604 |
+| **Masinissa-31M** | full fine-tune | **88.80%** | **0.8880** | **0.8831** | **0.9111** | **0.8697** |
 
-Linear Probe evaluates pure feature separability of the frozen pre-trained encoder (Mean Pooling → single `nn.Linear`). Full Fine-Tuning fits the encoder end-to-end with a `RobertaClassificationHead` (`Dense` → `Tanh` → `Dropout` → `Linear`).
+The **probe** is one `[hidden → 3]` layer over a frozen encoder, mean-pooled: it measures
+what pretraining already put in the representation, and nothing else can be credited for
+it. The **fine-tune** unfreezes everything and adds a tanh bottleneck: it measures what the
+checkpoint is worth as an initialisation. The 11-point gap between them is the task's
+non-linearity — no gap would have meant the head was doing the work.
+
+Both select their epoch on dev by macro F1 and score test once. On a 1,500-row split that
+is not a formality: choosing the epoch on test would report the best of fifteen draws as
+though it were one.
+
+```bash
+make modal-sentiment TASK=benchmark
+```
 
 ## Why this benchmark exists
 

dataset.stats.json

@@ -0,0 +1,27 @@
+{
+  "repo": "KabSentiment",
+  "files": [
+    {
+      "config": "default",
+      "split": "train",
+      "rows": 12000,
+      "path": "train.parquet",
+      "bytes": 672031
+    },
+    {
+      "config": "default",
+      "split": "dev",
+      "rows": 1500,
+      "path": "dev.parquet",
+      "bytes": 92688
+    },
+    {
+      "config": "default",
+      "split": "test",
+      "rows": 1500,
+      "path": "test.parquet",
+      "bytes": 93076
+    }
+  ],
+  "rows": 15000
+}

dev.parquet

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:f083510c996c720ba32e6c04e67e346524e81adae94c34ee4829c60395f58965
-size 56282
+oid sha256:eae46aa5f829ebe7216bb76234335efdac6188590a422c1d6db8bcc852bcabef
+size 92688

test.parquet

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:6585f07bd97d457bab755ac7d9b9bc9e3713513d2be3a0614765b7788fd89084
-size 56506
+oid sha256:ff329c0d994b1fb6bdaa019bb42c4e44ae3a5cb18d1e351639df6a86178d5953
+size 93076

train.parquet

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:84814a84c19e7252d4427b7c53549ae34fc4ad3674672e755e459d7c42048f09
-size 419011
+oid sha256:ae2fa2388d6c4fc7ce6e6130b0c98e48319c3fdd67667f7583ecc12d0678c8a5
+size 672031