license: apache-2.0
base_model: SargeDev/JEV-27B-VL-Uncensored
tags:
- decision-model
- system-one
- typed-decisions
- uncensored
- abliterated
- calibrated-probabilities
- noul
- choice
- score
- qwen3.8
- jev
- jev-1.13-teacher
- exl3
library_name: exllamav3
pipeline_tag: text-classification
datasets:- SargeDev/jev-distill-corpus-v3
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). The
quantization calibration prompts additionally sampled 384 prompts from SargeDev/solar-decisions-corpus-v4 (text only). - 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
EXL3 (exllamav3) quant of the vision flagship — the VL composite (vision tower + merger present,
MTP dropped per AutoTrust's convention) with the AutoTrust System-1 decision block merged into
huihui's abliterated Qwen3.8-27B backbone. Quantization: ~3.0 bpw routed weights, head layers at
6 bpw, embeddings at 16 bpw.
The decision adapter is already merged into these weights (backbone LoRA + lm_head LoRA) — do not
mount any additional decision adapter on top.
The decision machinery (head.safetensors, decision_head.json, judge_config.json,calibration.json, calibration_mergedfit.json, serve_decide.py, read_decision.py) ships in the
parent text repo SargeDev/JEV-27B-Uncensored;
the head projects the same 5120-dim text-tower hidden state, so the recipe applies to these weights too.
A quant-specific calibration battery for EXL3 precision is pending — treat the refit temperatures as
close but not yet validated at this precision. The published table below was measured on the bf16
weights before quantization. Vision decision battery results will be appended when finalized.
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) (bf16 weights, pre-quantization):
| 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).
Serving
Load with exllamav3 (>= 1.6), or serve via TabbyAPI (EXL3 support).
System 2 (chat/code/vision) runs through the normal lm_head with the bundled chat template.
For the calibrated System-1 decision engine use the parent repo's read_decision.py recipe
(bare-text prompt -> last-position final hidden state -> head projection -> per-kind temperature).
Vision decision battery results will be appended to this card when finalized.