sn92-generalist-erping-s50apk28

A LoRA fine-tune of perplexity-ai/pplx-decider-v1.2 at revision cce0f3f80033e9d89d80831d5281fd72124aacce, merged into the base weights, submitted to the REAX (Bittensor subnet 92) generalist decision track.

• Base model: perplexity-ai/pplx-decider-v1.2@cce0f3f80033e9d89d80831d5281fd72124aacce (Apache-2.0; itself derived from Qwen/Qwen3.8-27B, see NOTICE).
• Licence: Apache-2.0 (LICENSE is the base model's licence text; NOTICE carries the base notice forward).
• Format: unchanged Decider format. Backbone Qwen3_5Model safetensors shards with the same tensor names, dtypes and shapes as the base, plus readout.safetensors (255 x 5120 marker readout), decision_config.json and joint_config.json (joint protocol v3, <|fim_suffix|> marker readout, noncausal full attention). Run it exactly like the base model. The readout head is the fine-tuned head multiplied by 2.8 (stored in bf16) and decision_config.json temperature is 1.38235 x 2.8 = 3.87058, so the native Decider probabilities are unchanged while the REAX fixed-temperature readout (T = 2, last prompt position) gets the calibrated sharpness.
• This repository contains no code (the base repository's source/ package is not included, and config.json has no auto_map). Parameters: 26.09 B; weight files 48.6 GiB (bf16).

Method

• Low-rank adapters (LoRA) on the language-model projections only (attention q/k/v/o, gated-delta-net in_proj_qkv / in_proj_z / out_proj, MLP gate/up/down); the vision tower is not trained and its weights are byte-identical to the base. The readout head was trained jointly (initialised from the base readout).
• Objective: soft-target cross-entropy over each question's option logits at its marker position (the base model's own readout), with choice options shuffled per draw. Training used the base checkpoint's own inference path (noncausal attention over the full joint request), text only.
• Adapters were merged into the base weights in fp32 (W' = W + scale * B A) and stored in the base dtype (bf16).
• Model soup: the merged deltas of the adapters below were averaged with the listed weights (readouts averaged with the same weights).
  • adapter 1 (weight 0.5: mixture A, rank 32, alpha 64.0, targets attn_lin_mlp, step 869, tokens_seen 30006068)
  • adapter 2 (weight 0.5: mixture Bplus, rank 32, alpha 64.0, targets attn_lin_mlp, step 253, tokens_seen 6024281)

Training data

All rows are text decisions rendered in the base model's joint request format. Every source is used under its own licence; the attribution is below.

source	revision	split used	licence
decisionmodels/sn92-specialists	e2ff3bb0	train files only	CC-BY-4.0
tasksource/procedural-typed-decisions	d24af62a	train only	Apache-2.0
AmazonScience/massive 1.1 (8 non-English locales)	release tarball, sha256-pinned	train partition only	CC-BY-4.0
nvidia/Aegis-AI-Content-Safety-Dataset-2.0	d86bb8be	train only, human labels only	CC-BY-4.0
procedurally generated date/time and code/SQL decisions	ours	generated, gold by exact computation	Apache-2.0 (this repository)

Attribution: REAX / decisionmodels (sn92-specialists), tasksource / reasoning-core (procedural-typed-decisions), Amazon (MASSIVE), NVIDIA (Aegis 2.0).

Held out: no row of our evaluation lane (any split) was used for training. The sources of the lane's sealed confirmation panels were excluded at the dataset level, and every training row was checked against all lane rows (exact id, normalised text and 13-gram containment). Sealed confirmation rows were never used for training or calibration.

Intended use and limitations

• Intended for evaluation in the SN92 generalist track (classification-style decisions over ten areas, many languages) and for decision-model research.
• Text-only fine-tuning: image inputs behave like the base model at best and were not re-evaluated.
• Inherits the base model's limitations; probabilities can be miscalibrated on domains far from the training data.

Notice

pplx-decider-v1.2

Derived from Qwen/Qwen3.8-27B, revision 1d4bf0f2ff6012fd82039f2fa52739d0dd7c60c0, under Apache License 2.0.

Modified by ErPing (fine-tuned with LoRA, merged) on 2026-10-21.