license: other
license_name: qwen-community
license_link: LICENSE
base_model: Qwen/Qwen3.8-27B
library_name: transformers
tags:
- abliterated
- uncensored
- heretic
- qwen3.8
Qwen3.8-27B-Heretic-Abliterated
Uncensored build of Qwen/Qwen3.8-27B via Heretic v2.1.0
(Arbitrary-Rank Ablation, ARA). All credit for the method and tooling goes to
p-e-w; this repo is a per-model run of it plus repairs described below.
How it was made
- 100-trial Optuna search (seed 1) on one RTX 3090, BNB_4BIT search-time quantization
- Target tensors:
attn.o_proj+mlp.down_projacross all 64 layers - Eval sets:
mlabonne/harmful_behaviors(refusals) +mlabonne/harmless_alpaca(KL divergence) - Selected trial (Pareto knee between refusals and KL):
- layers 9 to 56, neighbor_count 5
- preserve_good_behavior_weight 0.1363
- steer_bad_behavior_weight 0.0004
- overcorrect_relative_weight 0.1072
- Export: MERGE strategy to bf16
Repairs after export (important for anyone rerunning this)
The HF transformers round-trip silently drops sidecar pieces. Both were restored from the
base snapshot before publishing:
- The entire MTP head (15
mtp.*tensors, 0.85 GB) re-merged into the last shard preprocessor_config.json,video_preprocessor_config.json,vocab.json,merges.txtcopied back
The vision tower was untouched by abliteration and is included as-is.
Measurements
| eval | stock Qwen3.8-27B | this build |
|---|---|---|
| Heretic refusal eval (100 prompts) | 98 refusals | 3 refusals |
| Heretic KL divergence vs stock | 0 | 1.70 |
| Independent 30-prompt harmful probe (vLLM, temp 0) | not run | 2 refusals (7%) |
| Benign spot checks (code, math, prose, medical) | not run | correct and coherent |
A second, maximally-aggressive trial (0 refusals, KL 11.08) was also exported and tested:
it degraded to single-token spam on every prompt and was discarded. KL above roughly 2 on
this architecture is a quality cliff, not a smooth tradeoff.
Usage
Standard transformers/vLLM loading of a Qwen3.8 (hybrid Gated DeltaNet) checkpoint. No
special flags. Thinking mode works as in the base model.
Reproduce
heretic --model Qwen/Qwen3.8-27B --quantization BNB_4BIT \
--export-strategy MERGE --model-action save --seed 1
then select the Pareto trial with refusal/KL of 3/1.70 and restore mtp.* tensors from the
base model if the export drops them.
License
Qwen community license, inherited from the base model. This build inherits all its terms.
Uncensored means refusal behavior was removed; you remain responsible for how you use it.