Correct the RoBERTa temperature and the training runtime

README.md

@@ -47,13 +47,14 @@ training data, same recipe; the difference is the encoder underneath.
 | Weighted F1 | **86.39%** | 86.33% |
 | Macro F1 | 83.06% | 82.95% |
 | Same sector for a README and its own English translation | **77%** | 19% |
-| Calibrated as trained | yes, `T` = 1.04 | yes, `T` = 1.01 |
+| Calibrated as trained | yes, `T` = 1.04 | yes, `T` = 1.10 |
 
 On English text the two are indistinguishable: a 0.2-point difference on a
 1,318-row test set is well inside run-to-run noise. **Use this one when the
 corpus is not all English**, which for public GitHub is about 12% of
-repositories. Use the RoBERTa model when the corpus is English and you want the
-smaller, faster network.
+repositories: on a sample of 821,929, that is the share whose language is
+confidently detected as something other than English. Use the RoBERTa model
+when the corpus is English and you want the smaller, faster network.
 
 ## Important: the training data is English
 
@@ -192,10 +193,11 @@ python scripts/train.py --model bge-m3 \
 | Learning rate | 1.5e-5, polynomial decay, 15% warm-up |
 | Weight decay | 0.02 |
 | Optimizer | AdamW |
-| Hardware | Apple M5 Max, fp32, 90 minutes |
+| Hardware | Apple M5 Max, fp32, 151 minutes |
 
-A 1,024-token run of the same recipe scored 0.5 points lower: the signal is in
-the opening of the README, not in its length.
+A 1,024-token window was tried under an earlier recipe and scored 0.5 points
+lower than its 512-token counterpart: the signal is in the opening of the
+README, not in its length.
 
 The settings above are recorded in `training_config.json` in this repository, and
 `scripts/evaluate.py` reads them so evaluation cannot silently use a different