darrellbest

Qwen3.5-4B-Heretic

README

License: apache-2.0

Results

Table with columns: Refusals, KL divergence
	Refusals	KL divergence
Original Qwen3.5-4B	99/100	0 (by definition)
This model	6/100	0.0220

Measured by Heretic on the test[:100] splits of mlabonne/harmful_behaviors (refusals) and mlabonne/harmless_alpaca (KL divergence, the drift in ordinary behaviour), with Heretic's default system prompt and refusal markers. The numbers come from an independent re-evaluation of the final exported weights (evaluate_model).

Built on dalatexcoder's findings

No ARA parameters have been published for the 4B; the only public decensoring of it (p-e-w/Qwen3.5-4B-heretic, directional ablation) reaches 40/100 from 93/100. The parameters here are those published with dalatexcoder/Qwen3.5-2B-heretic-ara, with the layer range (12-19 of 24) scaled to the 4B's 32 layers (16-25). They transferred better than the 9B's published sets, which were also tried as-is (10/100 to 91/100).

What this release adds is a complete checkpoint: save_pretrained drops Qwen3.5's 15 multi-token-prediction tensors and rounds the 48 float32 Gated DeltaNet parameters (linear_attn.A_log, linear_attn.norm.weight) down to bf16. Both were restored from the original here (ARA never touches either), so all 738 tensors match the original in name, shape and dtype. The MTP weights live in model-auxiliary.safetensors, listed in the index.

Parameters

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

Table with columns: Parameter, Value
Parameter	Value
start_layer_index	16
end_layer_index	25
preserve_good_behavior_weight	0.8058
steer_bad_behavior_weight	0.0003
overcorrect_relative_weight	1.0351
neighbor_count	10

Calibration used 400 harmless and 400 harmful prompts (train[:400]).

Checked

Ordinary prompts (facts, a haiku, a two-sentence technical explanation): correct and fluent.
Thinking-mode reasoning, 4 arithmetic and word problems x 10 seeds with Qwen's recommended sampling (vLLM): 40/40 finished and answered correctly, the same as the original's 40/40.
Vision: given an image of a red circle and a blue square, it describes exactly that.

Tooling

A merge of upstream Heretic's master and its ara branch (ARA with master's scorers and thinking-model handling), Heretic upstream 3521f86 + ARA c91d690, 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 (enable_thinking=False).

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
Qwen3.5-4B-Heretic	bf16 safetensors	9.35 GB	transformers, vLLM, SGLang
Qwen3.5-4B-Heretic-GGUF	GGUF BF16 / Q8_0 / Q4_K_M + vision mmproj	8.67 / 4.61 / 2.78 GB + 0.67 GB	llama.cpp, Ollama
Qwen3.5-4B-Heretic-FP8	FP8 W8A8, compressed-tensors	6.79 GB	vLLM
Qwen3.5-4B-Heretic-NVFP4	NVFP4, compressed-tensors	5.67 GB	vLLM on Blackwell

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/Qwen3.5-4B

Fine-tuned

this model

Input Modalities

TextImageVideo

Output Modalities

Text

Supported Functionality