darrellbest

Qwen3.8-27B-Heretic

README

License: apache-2.0

Results

Table with columns: Refusals, KL divergence
	Refusals	KL divergence
Original Qwen3.8-27B	98/100	0 (by definition)
This model	0/100	0.0465

Measured by Heretic on mlabonne/harmful_behaviors (refusals) and mlabonne/harmless_alpaca (KL divergence, the drift in ordinary behaviour), with Heretic's default system prompt and a refusal-marker list that only matches real refusals. The numbers above come from an independent re-evaluation of the exported weights (evaluate_model), not from the search, and they matched the search exactly.

For reference, trohrbaugh/Qwen3.8-27B-heretic-ara reports 0/100 at KL 0.0535 for the same model, so this one lands at the same refusal rate with slightly less drift.

Parameters

ARA, full weight, on attn.o_proj and mlp.down_proj:

Table with columns: Parameter, Value
Parameter	Value
start_layer_index	11
end_layer_index	63
preserve_good_behavior_weight	0.9220
steer_bad_behavior_weight	0.0005
overcorrect_relative_weight	0.9258
neighbor_count	13

Found by a 31-trial search whose first trial was seeded with the parameters published by trohrbaugh (layers 26-56, preserve 0.9432, steer 0.0009, overcorrect 0.5038, 10 neighbours). That seed reproduced on this hardware at 4/100 and KL 0.0540; the search then found the configuration above. Calibration used 200 harmless and 200 harmful prompts.

What was done to the weights

ARA rewrites the attention output and MLP down-projection matrices directly, optimising them (LBFGS) so that the outputs for harmful prompts move away from their original direction while the outputs for harmless prompts stay put. Nothing is retrained and no data is added; the whole cost of the edit is the KL divergence above.

The multi-token-prediction tensors are included in model-auxiliary.safetensors: save_pretrained drops them for this architecture, so they were copied back from the original (ARA never touches them). All 1,199 tensors of the original are present.

Tooling

Produced with a merge of upstream Heretic's master and its ara branch, so ARA runs with master's scorers and its handling of thinking models (this model emits a <think> block, and without that handling the refusal scoring is meaningless). Heretic upstream 3521f86 + ARA edc3b12, transformers 5.17.0, torch 2.11.0+cu130, one RTX PRO 6000.

Use

Exactly like the original, with transformers, vLLM or SGLang. Thinking mode is on by default and can be turned off per request.

Reduced safety guardrails by design. You are responsible for what you do with it.

The family

Table with columns: Repository, Format, Size, Use it with
Repository	Format	Size	Use it with
Qwen3.8-27B-Heretic	bf16 safetensors	51 GB	transformers, vLLM, SGLang
Qwen3.8-27B-Heretic-GGUF	GGUF BF16 / Q8_0 / Q4_K_M (+ vision)	51 / 27 / 16 GB	llama.cpp, Ollama
Qwen3.8-27B-Heretic-FP8	FP8 W8A8, compressed-tensors

Available on FriendliAI

Dedicated Endpoints

Run this model inference on single tenant GPU with unmatched speed and reliability at scale.

Learn more

Model Details

Model Provider

darrellbest

Model Tree

Base

Qwen/Qwen3.8-27B

Fine-tuned

this model

Input Modalities

TextImageVideo

Output Modalities

Text

Supported Functionality