nitrox/SA-SWE-32B-abliterated

SA-SWE-32B — Abliterated

Abliterated version of NovaSky-AI/SA-SWE-32B.

Base: 33B dense Qwen3 model fine-tuned for software engineering (SWE-bench). Extended thinking (<think>...</think>). BF16.

Abliteration

Performed with heretic — Optuna multi-objective optimization.

• Trials: 50
• Best trial: Trial #27, KL = 0.0003, ~2.1% refusals

Note on Quality

SA-SWE-32B is specialized for software engineering. On general-knowledge domains (security, chemistry, etc.) it may produce confident-sounding but inaccurate responses — inherent to the base model's specialization, not an artifact of abliteration.

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained(
    "nitrox/SA-SWE-32B-abliterated",
    device_map="auto",
    torch_dtype="bfloat16",
)
tokenizer = AutoTokenizer.from_pretrained("nitrox/SA-SWE-32B-abliterated")

messages = [{"role": "user", "content": "Your coding question here"}]
inputs = tokenizer.apply_chat_template(
    messages, add_generation_prompt=True, return_tensors="pt"
).to(model.device)
outputs = model.generate(inputs, max_new_tokens=4096)
print(tokenizer.decode(outputs[0][len(inputs[0]):], skip_special_tokens=True))

Disclaimer

Refusal mechanisms have been removed. Use responsibly and in accordance with applicable laws.