Update README.md

README.md

@@ -30,10 +30,10 @@ python filter_k5.py --in shard.parquet --out kept.parquet
 
 | Path | Size | Used by | What it is |
 |------|------|---------|------------|
-| `data/aoa.parquet` | 0.5 MB | Stage 1 | Age-of-Acquisition norms, 51,694 words |
-| `data/word_log_odds.parquet` | 1.9 MB | Stage 5 | Beyond-K–5 association scores, 129,719 words. `beyond_score` is `log2(rate_beyond / rate_k5)`; positive means characteristic of Beyond-K–5 material. |
-| `models/fasttext_grade.bin` | 57 MB | Stage 2 | fastText classifier, 4 classes |
-| `models/modernbert_grade/` | 299 MB | Stage 3 | ModernBERT classifier, 149.6M params, 4 classes |
+| `data/aoa.parquet` | 0.5 MB | Stage 1 | Age-of-Acquisition norms|
+| `data/word_log_odds.parquet` | 1.9 MB | Stage 5 | Beyond-K–5 association scores |
+| `models/fasttext_grade.bin` | 57 MB | Stage 2 | fastText classifier |
+| `models/modernbert_grade/` | 299 MB | Stage 3 | ModernBERT classifier |
 
 Both classifiers predict one of `K5`, `K8`, `K12`, `OOS`
 (`id2label = {0: "K5", 1: "K8", 2: "K12", 3: "OOS"}`); the filter retains
@@ -48,7 +48,7 @@ annotation of FineWeb-Edu would have been prohibitively expensive, which is
 what motivates the cascaded design: a cheap lexical stage, then fastText, then
 the ~50× more expensive ModernBERT.
 
-## Limitations
+## Intended Usage
 
 **Trained for web prose.** Labels come from FineWeb-Edu documents. On
 substantially different distributions — conversational text, worked solutions,