Repository navigation
feat: cascade attacks, base-model robustness runs and pinned requirements - #9
Merged
Merged
Conversation
…, and five re-attacked defences
… with result files (#2) * docs: pin requirements, add Makefile and CPU smoke path, align README with result files * docs: link the published adapters and add a download step
# Conflicts: # .gitignore
Add four untuned instruction models (Qwen2.5 0.5B/1.5B/3B, SmolLM2-1.7B) to the market robustness suite to separate the effect of task fine-tuning from model size, and a --markets flag so long runs split into resumable sittings. Result JSONs for attacks/variants/controls across the five markets included. Finding: task fine-tuning, not size, is what resists directional label-forcing; Neutral silencing resists neither.
study_cascade_nl.py: same TF-IDF features and training data, logistic vs MLP head, under the same greedy suppress/force attacks. The MLP gate is equally suppressible (~42% within 5 words) and only modestly harder to force, while forfeiting the certified word-count bound, which requires linearity.
# Conflicts: # run_robustness_markets.py
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
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
Brings the remaining robustness work into main: attacks on the confidence cascade with a certified word-count bound, untuned base-model runs, the linear vs non-linear cheap gate study, and the pinned requirements, Makefile and CPU smoke path.