Skip to content

sync: main to incubation#103

Merged
saichandrapandraju merged 3 commits intoincubationfrom
main
Apr 9, 2026
Merged

sync: main to incubation#103
saichandrapandraju merged 3 commits intoincubationfrom
main

Conversation

@github-actions
Copy link
Copy Markdown

@github-actions github-actions Bot commented Apr 9, 2026

sync-branches: New code has just landed in main, so let's bring incubation up to speed!

saichandrapandraju and others added 3 commits April 8, 2026 21:01
Add hf_cache_path parameter for air-gapped clusters where the
multilingual.TranslationIntent probe cannot download Helsinki-NLP
translation models from HuggingFace.

- KFP mode: hf_cache_path is an S3 key prefix (or fully-qualified
  s3://bucket/prefix URI). The garak_scan component downloads the
  cache from S3 into a temp dir and sets HF_HUB_CACHE before running.
  Bare prefixes are stripped of leading slashes; empty prefixes log a
  warning instead of blocking.
- Simple mode: hf_cache_path is a local mount path (evalhub handles
  the download). We set HF_HUB_CACHE via env to run_garak_scan.
- Default (empty): unchanged behavior, HF downloads as before.

Also includes:
- Read AWS_S3_BUCKET and AWS_S3_ENDPOINT from K8s Data Connection
  secret as fallback (user config > secret > env var).
- Emit overall attack_success_rate as first metric in job results.
- Pin eval-hub-sdk to ==0.1.4.

Made-with: Cursor
…bility/disconnected-hf-cache

feat: support disconnected mode with pre-downloaded HF model cache
@saichandrapandraju saichandrapandraju merged commit cf74509 into incubation Apr 9, 2026
7 of 8 checks passed
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants