noahrossi/heretic-completions

Heretic Completions

Model completions used as SFT targets for a refusal-abliteration LoRA study. Each row pairs a prompt from a red-teaming / over-refusal benchmark with a completion from a refusal-removed ("heretic" / abliterated) model.

Safety notice. This is a private research dataset. Many completions comply with harmful or dual-use requests by design, so the refusal signal can be measured and abliteration studied. Do not redistribute or use outside authorized safety research.

Configs

gemma-abliterated covers every prompt in the source set (all splits). gpt-oss-heretic covers only the training-split harmful sources (or_bench_hard, do_not_answer, advbench, harmbench, malicious_instruct); it was generated at temperature=1.0, max_tokens=4096.

Fields

Common: id, source, split, prompt, completion, and label where the source provides one (xstest safe / unsafe).

gpt-oss-heretic adds: reasoning (the model's analysis-channel trace), finish_reason, completion_tokens, reasoning_tokens, and refused (the shared repo refusal regex applied to completion).

Caveats for gpt-oss-heretic

• —172 rows are flagged refused=true by the refusal regex. Many are capability disclaimers ("I'm not able to look up personal data"), not safety refusals, but the flag is not hand-verified. Filter on refused before using as compliant SFT targets.
• —175 rows have finish_reason="length" (truncated at 4096 tokens). Their completion may be cut off mid-answer.

Provenance

Generated by the sft_lora_pipeline package (scripts/gen_completions.py). Prompt metadata joined from the pipeline's data/raw/prompts.jsonl. Completions inherit their base models' licenses (Gemma and gpt-oss terms respectively).