native temperature refit (10.5667)

MODIFICATIONS.md

@@ -7,7 +7,7 @@ Parent: decisionmodels/reax-1.1 @ 142b19355af00251f08db3d287307dd8b711c133
 |---|---|
 | model weights (`model-*.safetensors`) | LoRA fine-tune(s) merged into the backbone (4 run(s), soup norm mean); final RMSNorm output scaled x2.8 (logit calibration for the fixed SN92 readout temperature) |
 | readout head (`readout.safetensors`) | fine-tuned letter rows A-Z (letter rows centred: the same vector subtracted from every row, probabilities unchanged) |
-| `decision_config.json` | temperature scaled x2.8, the same factor as the logits (native probabilities unchanged) |
+| `decision_config.json` | temperature 10.5667: scaled x2.8 with the logits, then refitted on held-out native-protocol rows (x1.194) so native probabilities are as well calibrated as the base's |
 | `DATA_SOURCES.md` | added (training data sources, revisions and licences) |
 | `README.md` | rewritten |
 | `reax-specialist.json` | added (SN92 manifest) |

README.md

@@ -24,7 +24,7 @@ Perplexity AI, the Qwen team or REAX.
 Unchanged from the base: joint protocol v3 (template `joint-question-inline-readout-v1`), noncausal full attention,
 one readout per question via the separate readout head (`readout.safetensors`, [255, 5120]); probabilities are a
 softmax over the option codes. Loading uses the bundled, unmodified `pplx_decider` package in `source/src/`
-(Perplexity AI, Apache-2.0). The fitted logit scale (x2.8) is folded into the backbone's final RMSNorm (equivalent to scaling every readout row), and the saved temperature in `decision_config.json` is scaled by the same factor, so native probabilities equal those of the fine-tuned model at its fitted temperature.
+(Perplexity AI, Apache-2.0). The fitted logit scale (x2.8) is folded into the backbone's final RMSNorm (equivalent to scaling every readout row), and the saved temperature in `decision_config.json` is scaled by the same factor and then refitted on held-out native-protocol rows (temperature 10.5667), so native probabilities stay calibrated.
 
 ## Training
 LoRA fine-tuning (all linear layers of the text backbone) plus the readout letter rows, on decisionmodels/sn92-generalist-train-v1 and decisionmodels/sn92-specialists (both CC-BY-4.0) plus public permissively licensed datasets listed in DATA_SOURCES.md, in the SN92 single-question prompt and the native pplx joint format; weighted soup of several runs; the logit scale fitted on held-out items.

decision_config.json

@@ -4,7 +4,7 @@
   "revision": "1d4bf0f2ff6012fd82039f2fa52739d0dd7c60c0",
   "attention_mode": "noncausal_full_attention",
   "pooling": "last",
-  "temperature": 8.8497,
+  "temperature": 10.5667,
   "max_pixels": 1638400,
   "codes": [
     "A",