Qwen3.8-9B-Distill-heretic-ara-lora

This is a decensored version of empero-ai/Qwen3.8-9B-Distill, made using Heretic v1.2.0 with the Arbitrary-Rank Ablation (ARA) method (with row-norm preservation)

Heretic Command Arguments

The following command was utilised to generate this model: 'heretic --quantization bnb_4bit --model empero-ai/Qwen3.8-9B-Distill --batch-size 64 --use-ara-lora'

NOTE: This is a LORA. You will still need the base model in some form as it seems it wasn't automatically merged into the model despite select 'Merge LoRA into full model'. Whatever the case, the point of the matter is that you can choose to load in the LORA or not, but whatever you choose, you still need the actual model.

SIDE NOTE: Pretty much all of this README.md was generated automatically when I uploaded as well as the tags and what not.

Why?

Well, I actually completely missed the other heretic abliterations already done on this model and only noticed when I was going to upload. On the plus side-- it beats petruhonk/Qwen3.8-9B-Distill-uncensored-heretic in refusal counts

Abliteration parameters

Performance

---

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.

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

Quickstart

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

Sign up for the Empero newsletter at empero.org for releases, evals, and research notes.

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)