相同訓練的語言模型會在不同越獄提示下失守
Identically Trained Language Models Break on Different Jailbreaks
研究發現,相同數據配方及安全後訓練下,不同預訓練種子的語言模型仍會對不同越獄提示失守。在 25 種數據配方、共 150 個模型的比較中,同一預訓練運行的模型非拒答率相關係數為 0.984,換預訓練種子降至 0.811,換數據配方則為 0.779。
Published on Oct 25
Authors:
,
,
Abstract
Jailbreak vulnerability is primarily determined by pretraining randomness rather than data recipes, with pretraining seeds leaving identifiable fingerprints across prompts.
Jailbreak benchmarks test one checkpoint and treat which jailbreaks break it as a property of its pretraining data and safety training. We ask whether another pretraining run of the same data, given the same safety training, would break on the same jailbreaks. For each of 25 pretraining data recipes, we take 1B base models from 3 pretraining seeds and give each the same safety post-training with 2 post-training seeds, for 150 models in all. We then collect 20 responses from every model to a shared set of jailbreak prompts and correlate the per-prompt non-refusal rates between every pair of models. On average, two post-trained copies of the same pretraining run have per-prompt non-refusal rates that correlate at 0.984. A different pretraining seed lowers the correlation to 0.811, and a different data recipe lowers it only slightly further, to 0.779. These results indicate that which jailbreaks break a checkpoint depends much more on its pretraining run than on its data, a pattern that holds in all 25 data recipes and at 150M and 300M parameters. In practice, the best of 91 attack templates found on one checkpoint stays the best after a new pretraining seed in only 27% of model pairs, against 67% after a new post-training seed. Each pretraining run also leaves a fingerprint: across 3,211 prompts, a model's per-prompt non-refusal rates identify its pretraining run among all 75 runs every time, and a random subset of 50 prompts still does so 98% of the time.
View arXiv page View PDF GitHub 0 Add to collection
Models citing this paper 27
Updated 10 days ago • 11 • 1cerulean-works/castor-dolma1_7-no-reddit-1B
Updated 10 days ago • 8 • 1cerulean-works/castor-falcon-1B
Updated 10 days ago • 66cerulean-works/castor-dclm-baseline-1B
Updated 10 days ago • 24cerulean-works/castor-dclm-baseline-150M
Datasets citing this paper 1
cerulean-works/castor-labels
Viewer • Updated 10 days ago • 27.2M • 880
Spaces citing this paper 0
No Space linking this paper
Cite arxiv.org/abs/2610.34074 in a Space README.md to link it from this page.
Collections including this paper 1
來源:HuggingFace Daily Papers(社區熱門論文) · huggingface.co