MODIFICATIONS (Apache-2.0 sec. 4(b): notice that files were changed)
Product: REAX 1 (internal candidate id c2b) Version: c2b Date: 2026-10-20 Parent: perplexity-ai/pplx-decider-v1.2 @ cce0f3f80033e9d89d80831d5281fd72124aacce Root backbone: Qwen/Qwen3.8-27B @ 1d4bf0f2ff6012fd82039f2fa52739d0dd7c60c0
Changed files
| File | Change |
|---|---|
model weights (model-*.safetensors) |
four LoRA deltas merged into the backbone, each at weight 1/3: T-d (r16, hard-mined data), R1-a (r16, larger hard-mined data), T-e (r64, mixed generated/real data), saf-a (r16, 3,000 Aegis 2.0 train rows for safety classification + replay); all KL-to-base 0.5, lr 1e-4, 1 epoch |
readout head (readout.safetensors) |
(base readout + 1/3 * sum of the four fine-tuned readout deltas) x 1.41 x 1.4 (temperature folded into the head; the second factor 1.4 calibrates the fixed-T=2 REAX readout, 21 Oct 2026) |
decision_config.json |
temperature refit: 1.3823 -> 1.41 (NLL on our dev split, after the x1.41 head scaling), then x1.4 -> 1.9754 together with the head (native probabilities unchanged) |
README.md |
rewritten (new name, evaluation, attribution) |
source/src/pplx_decider/* |
unchanged (upstream loader code, Apache-2.0) |
LICENSE |
unchanged Apache-2.0 text |
NOTICE |
upstream NOTICE retained verbatim; modification lines added |
release-manifest.json |
removed (upstream hashes no longer apply); replaced by house-manifest.json |
Training summary
Data: ~4.4k-7.5k items per arm: base-hard-mined generated decision items (authors: OpenAI GPT-6 family, DeepSeek, GLM, Qwen, MiMo via APIs; verified by two other model families; contamination-guarded against all benchmark items we hold), plus 25 % easy replay, public SN92 specialist training rows (CC-BY-4.0) and 3,000 rows of the Aegis 2.0 TRAIN split (CC-BY-4.0, test-prompt overlap removed). Hardware: NVIDIA H200 (Verda, RunPod). Run ids: hm-r0-d, hm-r1-a, hm-r0-e, hm-saf-a (job sn92-house-model-20261020).