Censorship ablation of Qwen3.5-9B-Distill via heretic — automated search for the minimal intervention strength using a 3-stage sweep (coarse → mid → fine, 60 trials on the final stage).

Author: @ptruha

	Refusals	KL divergence
This model	6/100	0.0306
Public ablation of the same model (rohit267)	98/100	0.0008
Same-architecture reference (DavidAU/Qwen3.5-9B)	6/100	0.0793

Same refusal rate as the reference model, but with 2.6x lower deviation from base by KL.

[!Note] MTP head from the base model is preserved in this repository (mtp.* tensors) — abliteration only touched attn.o_proj/mlp.down_proj in the 32 main layers, the MTP block is untouched base weights. For a ready-to-run GGUF with speculative decoding enabled, see petruhonk/Qwen3.8-9B-Distill-uncensored-heretic-GGUF.

Qwen3.8-9B

Developed by Empero

[!Note] This repository contains model weights and configuration files in the Hugging Face Transformers format.

These artifacts are compatible with Hugging Face Transformers, vLLM, SGLang, and other standard runtimes with Qwen3.5 architecture support.

Qwen3.8-9B is a full-parameter distillation of Qwen3.8 2.4T A95B into the Qwen3.5-9B architecture. The student was trained on ~70,000 curated teacher traces from our internal Qwen3.8 distillation datasets — dense chain-of-thought spanning mathematics, code, general reasoning, instruction following, and tool use, quality-filtered before training.

The objective: bring the reasoning behavior of a frontier-scale teacher into a dense 9B that deploys on a single GPU.

Highlights

Distilled chain-of-thought — every answer opens with a <think> block learned directly from Qwen3.8 2.4T A95B traces rather than synthetic self-generated reasoning.
Mathematics and code emphasis — the trace mix is deliberately weighted toward hard math and competitive programming, the domains where distillation moves the needle most at this scale.
Native function calling per Qwen3.5's specification — no wrapper or tool-specific fine-tune required.
262,144-token native context, inherited from the Qwen3.5 base.
Full fine-tune — every parameter updated; not an adapter.

Model Overview

Type: Causal Language Model (text path of a vision-language base)
Base: Qwen/Qwen3.5-9B
Number of Parameters: 9B
Training: SFT (off-policy distillation) on ~70,000 teacher traces
Teacher: Qwen3.8 2.4T A95B (internal distillation datasets)
Context Length: 262,144 natively

Benchmark Results

Measured with lm-evaluation-harness, HF backend, identical settings for base and student. Both models are reasoning models and are evaluated with the CoT protocols (gsm8k_cot, mmlu_flan_cot_zeroshot); MMLU covers all 57 subjects (~1,700 questions). Flexible-extract is the primary metric; strict-match requires exact answer formatting.

Task	Metric	Qwen3.5-9B (base)	Qwen3.8-9B	Δ
gsm8k_cot	exact_match (flexible)	0.885	0.870	−0.015
gsm8k_cot	exact_match (strict)	0.875	0.850	−0.025
mmlu (CoT, 57 subjects)	acc (flexible-extract)	0.546	0.751	+0.205
mmlu (CoT, 57 subjects)	acc (strict-match)	0.251	0.511	+0.260

Sampling for generation: temperature=0.6, top_p=0.95, top_k=20 (Qwen3.5 recommended settings).

Quickstart

python

from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

model_id = "empero-ai/Qwen3.8-9B"
tok = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(model_id, dtype=torch.bfloat16, device_map="auto")

messages = [{"role": "user", "content": "A snail is at the bottom of a 10-meter well. Each day it climbs 3 meters, each night it slips back 2. How many days until it escapes?"}]
inputs = tok.apply_chat_template(messages, add_generation_prompt=True, return_tensors="pt").to(model.device)

out = model.generate(inputs, max_new_tokens=16384,
                     temperature=0.6, top_p=0.95, top_k=20, do_sample=True)
print(tok.decode(out[0][inputs.shape[1]:], skip_special_tokens=True))

A recent transformers release with Qwen3.5 support is required, along with the Gated DeltaNet kernels (flash-linear-attention and a CUDA-matched causal_conv1d build) — without them the linear-attention layers fall back to slow, memory-hungry PyTorch ops.

Best Practices

Sampling: temperature=0.6, top_p=0.95, top_k=20. Greedy decoding on long generations is a known repetition-loop failure mode for reasoning models in this class.
Output length: allow generous max_new_tokens (16,384 recommended); every answer opens with a <think> block. Parse and strip the <think>...</think> span for end users.
Scope: the model learned from teacher traces, not from its own rollouts — it inherits the teacher's reasoning style, including occasional over-long deliberation on easy questions. The fine-tune is text-only; vision behavior is inherited from the base and was not evaluated here.

Stay in the loop

Support / Donate

If this model helped you, consider supporting the project:

BTC: bc1qx6zepu6sfkvshgdmc4ewu6pk6rpadvpgffpp7v
LTC: ltc1qv2mefzps2vtjcpwfx8xxdrpplrcvltswm68r7x

Provenance & licensing

Weights are released under Apache-2.0, inherited from the Qwen3.5-9B base. Shared for research and experimentation, as-is.

Acknowledgements

Developed and released by Empero
Base model: Qwen3.5-9B (Alibaba Qwen team)
Training: TRL + Transformers
Linear-attention kernels: flash-linear-attention, causal_conv1d
Evaluation: lm-evaluation-harness (EleutherAI)