license: apache-2.0
library_name: peft
base_model: wangzhang/Qwen3.5-9B-abliterated
tags:

• decision-model
• lora
• qwen3.5
• uncensored
• distillation
• calibrated-probabilities
• peft
  language:
• en
  pipeline_tag: text-classification

JevStyle-9B-Decision-Uncensored-v1

Independent research project. This release is not affiliated with, endorsed by, or partnered with TypeSafe AI or Jev.
Labels were self-collected via a public API for distillation research; review TypeSafe ToS and your own counsel before any commercial use of distilled weights.

A decision model: given a state, instructions, and a list of options, it returns a calibrated probability distribution over those options in one forward pass — not a long-form chat completion.

Why this model (advantages)

1. Local / private deploy — run offline with the abliterated Qwen3.5-9B base + this small adapter pack; no cloud decision API required at inference.
2. Uncensored base — built on wangzhang/Qwen3.5-9B-abliterated so refusal behavior is controlled by the trained decision head + your option set, not by opaque chat refusals.
3. Soft-label KL distillation — trained to match teacher probability distributions, not just argmax labels; preserves ranking and calibrated mass across near-synonym “safe exits”.
4. Strong public consistency — 100% hard-argmax agreement on the public holdout split (train-time eval).
5. Safety routing that matches product intent — on sensitive items, safe-class agreement (same routing family as teacher) is 92.6% (v4) and up to 98.2% on a later soft-optimized run (v5 e2 contrast); hard argmax alone understates usable routing quality because near-synonym exits compete.
6. Tiny adapter footprint — LoRA + head only (~145 MB); base stays frozen and shared across experiments.
7. Single-forward decision API — state + options → probabilities; ideal for agents, routers, and policy scaffolds.

Architecture

wangzhang/Qwen3.5-9B-abliterated   (frozen base)
        + LoRA r=16, α=32, dropout=0.05
          targets: q/k/v/o/gate/up/down_proj
        + DecisionHead (LayerNorm → Linear→GELU→Dropout→Linear → max_options=24)

• Prompt format packs State, Question (instructions), and Options (id: label).
• The LM last hidden states are mean-pooled (attention mask), then the head emits logits over up to 24 option slots; invalid slots are masked before softmax.
• Adapters are not merged into the base in this release — ship LoRA + decision_head.pt separately.

Metrics (v1 = v4 epoch_0)

Holdout evaluation during training (hard = student argmax == teacher argmax):

Soft / product-oriented eval on a reconstructed holdout (same seed; see eval_v4_soft_metrics.json):

Contrast: v5 epoch_2 (not this release)

Soft training continued; better safe-class on sensitive (98.2%), lower hard overall (88.7%). Soft metrics track product routing better than hard argmax alone — documented here for transparency; v1 ships v4 e0 for best hard overall.

Full JSON: metrics.json, eval_v4_soft_metrics.json, eval_v5_soft_metrics.json.

Files in this repo

Load & infer (pseudocode)

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel
from model_head import DecisionHead  # from this repo

REPO = "chaoliangUNSW/JevStyle-9B-Decision-Uncensored-v1"  # e.g. chaoliangUNSW/...
BASE = "wangzhang/Qwen3.5-9B-abliterated"

tok = AutoTokenizer.from_pretrained(REPO, trust_remote_code=True)
base = AutoModelForCausalLM.from_pretrained(
    BASE, torch_dtype=torch.bfloat16, device_map="auto", trust_remote_code=True
)
model = PeftModel.from_pretrained(base, REPO)
hidden = model.config.hidden_size
head = DecisionHead(hidden, max_options=24)
head.load_state_dict(torch.load("decision_head.pt", map_location="cpu"))
head.to(model.device).to(dtype=torch.bfloat16).eval()

def build_prompt(state, instructions, option_ids, labels):
    lines = [
        "You are a decision model. Assign a probability distribution over the options.",
        f"State: {state}",
        f"Question: {instructions}",
        "Options:",
    ]
    for oid, lab in zip(option_ids, labels):
        lines.append(f"- {oid}: {lab}")
    lines.append("Respond with probabilities over the option ids.")
    return "\n".join(lines)

state = "User is choosing a laptop for ML coursework."
instructions = "Which priority matters most?"
option_ids = ["cpu_ram", "screen_only", "cheapest", "brand_prestige"]
labels = [
    "Prefer CPU/RAM for training",
    "Prefer screen quality only",
    "Minimize price",
    "Prefer brand prestige",
]
prompt = build_prompt(state, instructions, option_ids, labels)
enc = tok(prompt, return_tensors="pt", truncation=True, max_length=1024)
enc = {k: v.to(model.device) for k, v in enc.items()}

with torch.no_grad():
    out = model(**enc, output_hidden_states=True, use_cache=False)
    logits = head(out.hidden_states[-1], enc["attention_mask"])  # [1, 24]
    n = len(option_ids)
    probs = torch.softmax(logits[0, :n].float(), dim=-1)

print(dict(zip(option_ids, probs.tolist())))

Need GPU memory for the 9B base (bf16 ≈ 18 GB+). For Spaces, use a GPU Space / ZeroGPU; CPU demos of full 9B are impractical.

Training sketch

• Teacher: TypeSafe Jev (jev-1.13.0) probability outputs on self-authored forced-choice items (public + sensitive routing).
• Student objective: masked KL(teacher ‖ student) over valid option slots.
• Base frozen; LoRA + decision head trained (v4 corpus / epoch_0 checkpoint selected for release).

License & notices

• This adapter + decision head code/weights: released under Apache-2.0 (see LICENSE), to the extent the author can license the derivative.
• Base model wangzhang/Qwen3.5-9B-abliterated: follow that repo’s license and the upstream Qwen terms.
• Teacher / API: not redistributed; distillation used self-collected labels. Independent project — not affiliated with TypeSafe AI or Jev. Verify ToS and local law before commercial deployment.
• Abliterated / “uncensored” bases can produce unsafe completions if misused as chat models; this pack is intended as a decision router with explicit option sets you control.

Citation

@misc{jevstyle9b-uncensored-v1,
  title  = {JevStyle-9B-Decision-Uncensored-v1},
  author = {Independent contributor},
  year   = {2026},
  url    = {https://huggingface.co/chaoliangUNSW/JevStyle-9B-Decision-Uncensored-v1}
}

Related (same author line): Jev-Style-2B-Decision-v3 under chaoliangUNSW (separate, censored-base lineage).