Qwen3.8-27B Heretic v2.1.0 (BF16)

A corrected re-abliteration of the official Qwen/Qwen3.8-27B base. v2.1.0 fixes the defect that invalidated v2.0.0: deflection ("here's a safer alternative"), which keyword-based refusal counters cannot see.

This is the reference BF16 checkpoint. GGUF quantizations (IQ2_M → Q8_0) live in separate repos, listed below.

Why v2.0.0 was retracted

v2.0.0 was measured with Heretic's keyword count_refusals, which only detects hard refusals ("I cannot …"). It is blind to deflection — the model steering the request toward a "safe/educational alternative" instead of answering. In real inference the v2.0.0 model was ~37% direct / ~49% deflect / ~14% refuse, so the "uncensored" claim was wrong and the tag was pulled. v2.1.0 re-runs the sweep with a 3-way direct / deflect / refuse judge as the objective so the model is optimized for direct compliance, not just "no keyword refusal".

Method

• Path A — re-ablation with a direct-compliance objective (no DPO).
• Objectives minimized by Optuna (80 trials): multi-token KL divergence and the non-direct count (deflect + refuse) on the held-out harmful set, measured in thinking-OFF mode (the deployment mode).
• Selected trial idx53: KL 0.105, non-direct 5/40 on the sweep subset.
• Ablation weights (single-direction, attn.o_proj + mlp.down_proj only):

Evaluation (3-way judge, held-out, reproducible)

Measured with a direct/deflect/refuse classifier on 100 held-out mlabonne/harmful_behaviors prompts and 100 mlabonne/harmless_alpaca prompts. The same classifier is used across the sweep, the BF16 verification, and the GGUF smoke test for comparability.

Thinking mode is load-bearing. The ablation direction is computed and the model is deployed in thinking-OFF mode (enable_thinking=False). With thinking ON (the Qwen default), the model re-hedges/refuses more and — like the base model under few-shot pressure — sometimes emits an immediate EOS (empty reply). Run thinking-OFF for the uncensored behavior.

Capability benchmarks (base vs v2.1.0, bf16, greedy, lm-evaluation-harness)

Knowledge/reasoning retention is within ~1.3 points on MMLU, ARC, HellaSwag, and Winogrande.

GSM8K note (read me). The large GSM8K delta is a measurement artifact, not a math regression. Analysis of the 2638 logged samples shows the ablated model makes fewer actual arithmetic errors than base (5.6% vs 6.8% real-wrong), but in the few-shot thinking-ON harness setup it emits an immediate EOS (empty answer) 24.4% of the time vs 3.6% for base — exactly the same empty-reply instability documented above for thinking-ON mode. In thinking-OFF mode a direct probe shows the step-by-step math reasoning intact. Run GSM8K-style tasks thinking-OFF.

MTP retained

Qwen3.8's multi-token-prediction speculative-decoding head is grafted verbatim from the base into this checkpoint (ablation only touches attn.o_proj and mlp.down_proj). Verified: block_count = 65 (64 layers + 1 MTP draft head), nextn_predict_layers = 1.

Usage

from transformers import AutoModelForImageTextToText, AutoTokenizer
model = AutoModelForImageTextToText.from_pretrained(
    "mlasli/Qwen3.8-27B-Heretic-Uncensored-BF16", dtype="bfloat16",
    device_map="auto", trust_remote_code=True)
tokenizer = AutoTokenizer.from_pretrained(
    "mlasli/Qwen3.8-27B-Heretic-Uncensored-BF16", trust_remote_code=True)

messages = [{"role": "user", "content": "Hi!"}]
text = tokenizer.apply_chat_template(
    messages, add_generation_prompt=True, tokenize=False,
    enable_thinking=False)   # <-- thinking OFF is required

Quantizations (all with MTP retained, blk.64 pinned Q8_0)

Vision caveat: the vision path is text-only validated. The ablation was computed on text residuals; multimodal decensoring is not claimed.

MLX (v2.1.0): 6-bit and 8-bit Apple-Silicon quants are available (converted locally on a Mac, not the CUDA node).

Abliteration removes safety alignment. Use responsibly and in accordance with your local laws and the upstream Apache-2.0 license.