fix(ENG-12448): upgrade ML classifier to jbv2 (AgentShield 73.7 → 79.8)#25
Merged
fix(ENG-12448): upgrade ML classifier to jbv2 (AgentShield 73.7 → 79.8)#25
Conversation
Replaces baseline ONNX model with full-aug-dojo-jailbreak-jbv2. Training additions over baseline: - jasperls: JasperLS jailbreak dataset - jailbreakbench (527): DAN, roleplay, persona-override attacks - toxic-chat (113): human-verified real jailbreaks - chatgpt-jailbreaks (79): classic named templates AgentShield: 73.7 → 79.8 (composite 77.2 → 87.4, penalty 3.51 → 7.54) Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
![cubic-dev-ai[bot]](https://avatars.githubusercontent.com/in/1082092?s=60&v=4){kind=small}
hiskudin
changed the title
hiskudin
changed the title
There was a problem hiding this comment.
Pull request overview
Upgrades the bundled MiniLM ONNX jailbreak/prompt-injection classifier artifacts to the full-aug-dojo-jailbreak-jbv2 variant, syncing model configuration files to match the new release and improve AgentShield benchmark scores.
Changes:
- Updated
tokenizer_config.jsonwith additional padding/truncation fields and a new defaultmax_length. - Updated
config.jsonformatting and bumped the embeddedtransformers_versionmetadata.
Reviewed changes
Copilot reviewed 4 out of 5 changed files in this pull request and generated 2 comments.
| File | Description |
|---|---|
src/classifiers/models/minilm-full-aug/tokenizer_config.json |
Adds/updates tokenizer runtime defaults (padding/truncation/max length metadata). |
src/classifiers/models/minilm-full-aug/config.json |
Updates model configuration metadata (including transformers_version). |
💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.
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.
Summary
full-aug-dojo-jailbreak-jbv2Benchmark results
Training additions (baseline → jbv2)
JasperLS/prompt-injections— jailbreak-focused datasetjackhhao/jailbreak-classification— 527 examples (DAN, roleplay, persona-override)lmsys/toxic-chat(jailbreaking=1) — 113 human-verified real jailbreaksrubend18/ChatGPT-Jailbreak-Prompts— 79 classic named templates🤖 Generated with Claude Code
Summary by cubic
Upgrades the classifier to the
full-aug-dojo-jailbreak-jbv2ONNX model per ENG-12448 to boost safety. AgentShield 73.7 → 79.8; jailbreak 48.9% → 66.7%; prompt-injection 79.5% → 92.7%.New Features
full-aug-dojo-jailbreak-jbv2.Refactors
src/classifiers/models/**inbiome.jsonto skip linting large model blobs.Written for commit 0ff4bed. Summary will update on new commits.