Update README.md

README.md

@@ -77,6 +77,7 @@ print(clf({"text": "The cat is sitting on top of the bookshelf.", "text_pair": "
 - **Split:** group-aware 70/15/15 by speaker session (no session spans splits)
 - **Hyperparameters:** 2 epochs · lr 2e-5 · batch 16 · weight decay 0.01 · warmup 0.1 ·
   max_length 128 · fp16 · seed 42 · best checkpoint by F1
+- **Class imbalance:** no class weighting or resampling. The dictionary gate already raises the positive rate from ~4% of all words to ~30% of candidates, and checkpoint selection used minority-class F1 rather than accuracy.
 - **Framework:** 🤗 Transformers
 
 ## Evaluation (held-out test set)
@@ -85,18 +86,17 @@ Reported for two views: **dictionary candidates only** (the meaningful view —
 dictionary flags as plausibly spatial) and **overall** (every word, dominated by trivially
 non-spatial tokens).
 
-### Candidates only — 799 words (570 not-spatial, 229 spatial)
+### Candidates only — 622 words (393 not-spatial, 229 spatial)
 
 | class | precision | recall | F1 | support |
 |---|---|---|---|---|
-| not_spatial | 0.958 | 0.925 | 0.941 | 570 |
+| not_spatial | 0.938 | 0.891 | 0.914 | 393 |
 | spatial | 0.827 | 0.900 | 0.862 | 229 |
 
-**Overall accuracy: 0.917** (733 of 799 words correct) · macro-F1 0.901 · Cohen's κ 0.803
+**Overall accuracy: 0.894** (556 of 622 words correct) · macro-F1 0.888 · Cohen's κ 0.776
 
 Reading it per class: the model catches **90.0%** of truly-spatial words (recall) at **82.7%**
-precision; for non-spatial words it's 92.5% recall at 95.8% precision. (In clinical terms:
-sensitivity 0.900, specificity 0.925, PPV 0.827, NPV 0.958.)
+precision; for non-spatial words it's 89.1% recall at 93.8% precision. (sensitivity 0.900, specificity 0.891, PPV 0.827, NPV 0.938.)
 
 ### Overall — every word, 5,207 tokens
 
@@ -104,13 +104,13 @@ accuracy **0.987** · spatial-F1 **0.862** · Cohen's κ **0.855**
 
 > The two views share the *same* spatial predictions (229 spatial words, same 206 caught). Only
 > the non-spatial pool differs, which is why "overall" accuracy looks higher — it's padded with
-> ~4,400 easy non-candidate words the model trivially gets right. Judge the model by the
+> ~4,600 easy non-candidate words the model trivially gets right. Judge the model by the
 > **candidates-only** view.
 
 ## Calibration
 
 Raw probabilities are over-confident, so a post-hoc **temperature scaling** factor
-(**T = 1.816**, fit on the validation candidates) rescales them into a calibrated `P(spatial)`
+(**T = 1.887**, fit on the validation candidates) rescales them into a calibrated `P(spatial)`
 you can read literally. Temperature scaling is monotonic, so the hard 0/1 decision is unchanged.
 See section 7.5 of the repo.