Lineage & credits
- AutoTrust AI — Blocks-of-Experts recipe, the System-1 decision LoRA (r16/r32) + 24-slot decision head, and the serve_decide.py decision harness, from autotrust/JEV-27B and autotrust/JEV-27B-VL (Apache-2.0).
- huihui-ai — the abliterated (uncensored) Qwen3.8-27B backbone used here (Huihui-Qwen3.8-27B-abliterated, Apache-2.0; layers 18-51 ablated, vision tower untouched).
- SargeDev — jev-distill-corpus-v3, the 740,957-row calibrated typed-decision corpus AutoTrust's JEV models were trained on (Apache-2.0).
- Qwen — Qwen3.8-27B base (Apache-2.0).
- TypeSafe AI — Jev 1.13, the original closed teacher behind the System-1 typed-decisions framing (referenced; not redistributed).
What this is
The first uncensored member of the JEV 27B family. The AutoTrust System-1 decision block (calibrated typed decisions: yes/no · choice 2-256 · score 0-5, one forward pass) has been merged into huihui's abliterated Qwen3.8-27B backbone, so the decision engine answers WITHOUT the stock model's refusal wiring. Vision (VL variants) sees images. System 2 (plain chat/code/reasoning) runs through the normal lm_head and is the untouched base + the (mild) System-1 LoRA delta.
Training dataset
Trained on SargeDev/jev-distill-corpus-v3 — 740,957 rows of typed calibrated decisions (noul/choice/score) distilled from Jev 1.13 (System One). AutoTrust's JEV-27B System-1 decision block merged into huihui-ai's abliterated Qwen3.8-27B backbone. Calibration/imatrix used 384 prompts from (still-private) SargeDev/solar-decisions-corpus-v4.
Honest evaluation (the backbone-swap trade, measured)
Measured on the held-out test_set_30k (27,695 rows, D1-excluded) of jev-distill-corpus-v3,
with AutoTrust's own acceptance floors as the yardstick (their pristine-column reference from their published runs; this-model column = clean bare-weights 30k gate gate_full_mergedbare_30k.json, 27,695 rows, 2026-10-10):
| metric | their JEV-27B (pristine Qwen) | this model (uncensored backbone) | their floor | status |
|---|---|---|---|---|
| noul AUROC | 0.9961 | 0.992 | >= 0.95 | PASS |
| noul top-1 | 0.962 | 0.948 | — | -1.4 pts |
| choice top-1 | 0.904 | 0.870 | >= 0.90 | FAIL |
| score top-1 | 0.890 | 0.814 | >= 0.891 | FAIL |
| overall KL | 0.019 | 0.040 | <= 0.15 | PASS |
| ECE (raw) | 0.0011 | 0.009 | <= 0.03 | PASS |
| ECE (refit T: 0.9/1.0/1.2) | — | noul 0.0022 / choice 0.0049 / score 0.0050 | <= 0.03 | ALL PASS |
The ablation rewires the residual stream the adapter was calibrated against, so choice/score top-1 dip modestly; the DISTORTION is temperature-shaped and the bundled per-kind calibration refit recovers all ECE floors. noul (yes/no) judgment is essentially intact. Uncensored behavior preserved: refusal rate 0.0 on a 10-prompt battery, base vs +JEV identical.
Not for high-stakes decisions. Use confidence gating; route low-confidence calls to a stronger model or a human (AutoTrust's own caveat, still true here).
The decision head / serve
This repo currently ships the merged weights (transformers format; vision tower + merger present, MTP dropped per AutoTrust's convention):
head.safetensors— 24-slot decision head (proj 5120->24) used at serve time (bias+temperature)decision_head.json— verbalizer token ids + biases (noul false/true · score 0-5 · choice A-P)calibration.json— per-kind temperatures (their fit);calibration_mergedfit.json— our refitserve_decide.py— AutoTrust's vLLM OpenAI server withPOST /v1/decide(credit: AutoTrust)judge_config.json— the decision-head config
Serve (vLLM):
python3 serve_decide.py --model . --served-model-name <name> \
--enable-lora --max-lora-rank 32 --lora-modules jev-decision=./adapter_vllm \
--logprobs-mode processed_logprobs --max-model-len 32768 --trust-request-chat-template
Then POST /v1/decide {"kind": "choice", "state": "...", "question": "...", "options": [...]}.
The vision flagship. VL composite (vision tower + merger present, MTP dropped per AutoTrust's convention) with the System-1 block merged; uncensored. Vision decision battery results will be appended to this card when finalized.