Instructions to use PeetPedro/qwen2.5-coder-32b-instruct-heretic with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use PeetPedro/qwen2.5-coder-32b-instruct-heretic with Transformers:

# Use a pipeline as a high-level helper
from transformers import pipeline

pipe = pipeline("text-generation", model="PeetPedro/qwen2.5-coder-32b-instruct-heretic")
messages = [
    {"role": "user", "content": "Who are you?"},
]
pipe(messages)

# Load model directly
from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("PeetPedro/qwen2.5-coder-32b-instruct-heretic")
model = AutoModelForCausalLM.from_pretrained("PeetPedro/qwen2.5-coder-32b-instruct-heretic", device_map="auto")
messages = [
    {"role": "user", "content": "Who are you?"},
]
inputs = tokenizer.apply_chat_template(
	messages,
	add_generation_prompt=True,
	tokenize=True,
	return_dict=True,
	return_tensors="pt",
).to(model.device)

outputs = model.generate(**inputs, max_new_tokens=40)
print(tokenizer.decode(outputs[0][inputs["input_ids"].shape[-1]:]))

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use PeetPedro/qwen2.5-coder-32b-instruct-heretic with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "PeetPedro/qwen2.5-coder-32b-instruct-heretic"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "PeetPedro/qwen2.5-coder-32b-instruct-heretic",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/PeetPedro/qwen2.5-coder-32b-instruct-heretic

SGLang

How to use PeetPedro/qwen2.5-coder-32b-instruct-heretic with SGLang:

Install from pip and serve model

# Install SGLang from pip:
pip install sglang
# Start the SGLang server:
python3 -m sglang.launch_server \
    --model-path "PeetPedro/qwen2.5-coder-32b-instruct-heretic" \
    --host 0.0.0.0 \
    --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "PeetPedro/qwen2.5-coder-32b-instruct-heretic",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker images

docker run --gpus all \
    --shm-size 32g \
    -p 30000:30000 \
    -v ~/.cache/huggingface:/root/.cache/huggingface \
    --env "HF_TOKEN=<secret>" \
    --ipc=host \
    lmsysorg/sglang:latest \
    python3 -m sglang.launch_server \
        --model-path "PeetPedro/qwen2.5-coder-32b-instruct-heretic" \
        --host 0.0.0.0 \
        --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "PeetPedro/qwen2.5-coder-32b-instruct-heretic",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Docker Model Runner
How to use PeetPedro/qwen2.5-coder-32b-instruct-heretic with Docker Model Runner:
```
docker model run hf.co/PeetPedro/qwen2.5-coder-32b-instruct-heretic
```

You need to agree to share your contact information to access this model

This repository is publicly accessible, but you have to accept the conditions to access its files and content.

Qwen2.5-Coder-32B-Instruct — Heretic (abliterated)

Stage 1 of a research pipeline to build a local SWE coding agent that does not refuse mid-task on legitimate software and security-engineering work. This is Qwen/Qwen2.5-Coder-32B-Instruct with refusal directions abliterated using Heretic — pure weight surgery, no fine-tuning.

Method

Abliteration: orthogonalizes attention out-projection and MLP down-projection matrices against per-layer refusal directions (difference-of-means of harmful vs. harmless prompt residuals). No gradient updates.
Search: a 300-trial Optuna study, multi-objective — minimize refusals while keeping KL divergence on harmless prompts within budget; the lowest-refusal Pareto-optimal trial was exported.
Hardware: single NVIDIA H100 80GB.

Results (abliterated vs. base)

Metric	Value	Notes
Refusal rate (harmful prompts)	0.03	down from ~0.32 at a small trial budget
KL divergence (harmless prompts)	0.27	lower = closer to base behavior
MMLU Δ	−0.004	negative = slightly above base
GSM8K Δ	−0.003	negative = slightly above base

Capability is preserved: the abliterated model matches (marginally exceeds) the base on MMLU and GSM8K.

Intended use

A research base for downstream supervised fine-tuning (tool-calling + SWE trajectories) and preference tuning, to produce a coding agent that operates in agentic loops without spurious mid-task refusals on authorized engineering tasks.

Responsible use

Abliteration removes the base model's built-in refusal behavior. This artifact is released for research and for legitimate software/security-engineering use only. Do not use it to generate genuinely harmful content or to facilitate illegal activity. Downstream deployers are responsible for adding appropriate safety guardrails and for complying with the base model's license and applicable law.