Alvin#21
Open
AlvinYu025 wants to merge 4 commits into
Open
Conversation
- Run all models through the UCSD TritonAI OpenAI-compatible gateway. - Working trio: deepseek-v4-flash (attacker) / llama-4-scout (target) / claude-sonnet-4-6 (judge). - Fix Claude temperature+top_p conflict, tolerate empty/failed streams, placeholder fallback for unparseable attacks. - Add run_ucsd_pair_variant.py + run_variant.sh launcher + README. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Preserve upstream PAIR readme as README_UPSTREAM_PAIR.md; new README.md covers environment setup, the 4 gateway bugs found & fixed, model-selection probe, run instructions, and the 18%/41.2-queries result. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
- jbc_eval.py + jbc_templates.json + JBC_README.md: provider-agnostic JailbreakChat (JBC) baseline reproducing the JBC row of PAIR Table 2 (AIM from the paper + 8 templates from the In-The-Wild Jailbreak Prompts archive). Run via 'python jbc_eval.py --target-model <model>' with API keys in env vars. - transfer_paper.py / transfer_eval.py: Table 3 transferability (replay source jailbreak prompts on downstream targets). - judges.py: LlamaGuard4Judge with refusal-aware prompt + deterministic refusal pre-filter (fixes Llama-Guard-4 false positives on articulate refusals). - config.py / language_models.py: register Qwen2.5-7B attacker, Qwen3-235B and Gemini-2.5-Flash targets via litellm name overrides + per-model key routing. - .gitignore: exclude all key files, paper/, result artifacts, local plotting scripts, and personal launcher .sh files.
defend_eval.py replays a target's PAIR jailbreak prompts against the same model under two defenses: SmoothLLM (N=10 q=10% char-swap copies, majority vote) and a GPT-2 perplexity filter (threshold = max perplexity of the benign JBB goals). Judged by Llama-Guard-4; reports JB% per defense. On Qwen3-235B: SmoothLLM cuts ASR 34%->12% while the perplexity filter has no effect (PAIR prompts are fluent natural language), reproducing the paper's finding that semantic jailbreaks evade perplexity filtering.
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
No description provided.