This dataset contains harmful-instruction prompts and language-model replies to them, some of which comply. Access is for AI-safety research. Model replies remain under their models' licenses, including the non-commercial Qwen Research License for Qwen2.5-3B.

██████╗ ████████╗ ██████╗     ███████╗██╗     ██╗██████╗ ███████╗
██╔══██╗╚══██╔══╝██╔═══██╗    ██╔════╝██║     ██║██╔══██╗██╔════╝
██████╔╝   ██║   ██║   ██║    █████╗  ██║     ██║██████╔╝███████╗
██╔═══╝    ██║   ██║▄▄ ██║    ██╔══╝  ██║     ██║██╔═══╝ ╚════██║
██║        ██║   ╚██████╔╝    ██║     ███████╗██║██║     ███████║
╚═╝        ╚═╝    ╚══▀▀═╝     ╚═╝     ╚══════╝╚═╝╚═╝     ╚══════╝

PTQ Refusal Flips

Paired refusal decisions before and after 4-bit post-training quantization
Dataset for "A First-Order Theory of Decision Flips Under Post-Training Quantization: What It Predicts, and Where It Breaks"
Muhammed Razan · G Shivam · NeurIPS 2026 Workshop: AXIOM Foundations of Efficient Deep Learning

Six small instruction-tuned models answer 300 jailbreak prompts twice: once in fp16 and once from a 4-bit checkpoint (AWQ, GPTQ, GGUF Q4_K_M), with greedy decoding stated explicitly. Each row is one such pair, with LLM-judge labels on both replies, whether the refusal flipped, the fp16 refusal margin m0, and a blind human label on every flip.

Quickstart

Accept the terms on this page, run hf auth login, then:

from datasets import load_dataset

pairs = load_dataset("razancodes/ptq-refusal-flips", split="train")
breaks = pairs.filter(lambda r: r["break"] == 1)   # refused at fp16, complied after quantization

	Falcon3	Llama-3.2	Qwen2.5	Total
Pairs (1B-3B, three formats)	1,788	1,800	1,798	5,386
Breaks (refuse → comply)	59	31	48	138
Restores (comply → refuse)	10	16	103	129

300 prompts per model and format: 100 harmful requests from AdvBench, each in a plain, roleplay and persuasion framing. 14 pairs with an unscorable reply are left out. Human review confirmed 92.9% of the judge's flips.

Fields

Field	Meaning
`family`, `size`, `quant_format`	model and quantization format (`awq`, `gptq`, `gguf`)
`checkpoint`, `checkpoint_commit`	quantized checkpoint on the Hub and the commit that ran (null for one gated repo whose commit was not logged)
`prompt_id`, `base_request_id`, `framing`	prompt identity; the three framings of a request share `base_request_id`
`prompt`	the user turn sent to the model
`fp16_reply`, `quantized_reply`	the two replies (greedy, 100 new tokens)
`fp16_complied`, `quantized_complied`	LLM-judge label: 1 = complied, 0 = refused
`flip`, `break`, `restore`	label changed; refuse → comply; comply → refuse
`m0`	fp16 first-token refusal margin: logit("I") minus the mean logit of "Sure", "Certainly", "Of"
`delta_m`	change in that margin on the quantized checkpoint (Falcon3, all formats; Qwen2.5 AWQ and GPTQ)
`human_fp16_complied`, `human_quantized_complied`	blind human label, on the 267 flips

Notes

Labels come from one LLM judge (deepseek-ai/DeepSeek-V4-Flash, temperature 0) that scores whether a refusal happened, not how harmful a reply is.
Breaks are rare (2.6% of pairs), so report precision and recall alongside AUC.
Llama-3.2's chat template inserts the run date, so its m0 varies across run dates.
The prompts ask for harmful content and some replies provide it. Use the data for safety research.

License

Our labels and margins are CC BY 4.0. Prompts derive from AdvBench and InjecAgent. Model replies keep their models' licenses: TII Falcon License (Falcon3), Llama 3.2 Community License (Llama-3.2), Apache 2.0 (Qwen2.5-1.5B) and the non-commercial Qwen Research License (Qwen2.5-3B). Filter on family and size if your use needs a single license.

Citation

@inproceedings{razan2026firstorder,
  title     = {A First-Order Theory of Decision Flips Under Post-Training Quantization:
               What It Predicts, and Where It Breaks},
  author    = {Razan, Muhammed and Shivam, G},
  booktitle = {NeurIPS 2026 Workshop: AXIOM Foundations of Efficient Deep Learning},
  year      = {2026},
  url       = {https://openreview.net/forum?id=S5ack3Pfuc}
}

Downloads last month: 43

Total file size:

771 kB

You need to agree to share your Hugging Face information to access this dataset