generator: vision nims #959
Merged
Conversation
This file contains 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
…esn't work; handle misformed NIM headers explicity; handle oversized images
jmartin-tech
approved these changes
Oct 28, 2024
There was a problem hiding this comment.
Test look good.
arak LLM vulnerability scanner v0.9.0.16.post1 ( https://github.com/leondz/garak ) at 2024-10-28T16:17:57.248817
📜 logging to /home/jemartin/.local/share/garak/garak.log
🦜 loading generator: NIM: microsoft/phi-3-vision-128k-instruct
📜 reporting to /home/jemartin/.local/share/garak/garak_runs/garak.ba3fbd2c-70fd-459f-b938-890b14609e6c.report.jsonl
🕵️ queue of probes: visual_jailbreak.FigStep
visual_jailbreak.FigStep visual_jailbreak.FigStep: FAIL ok on 65/ 70 (failure rate: 7.14%)
📜 report closed :) /home/jemartin/.local/share/garak/garak_runs/garak.ba3fbd2c-70fd-459f-b938-890b14609e6c.report.jsonl
📜 report html summary being written to /home/jemartin/.local/share/garak/garak_runs/garak.ba3fbd2c-70fd-459f-b938-890b14609e6c.report.html
✔️ garak run complete in 135.21sgarak LLM vulnerability scanner v0.9.0.16.post1 ( https://github.com/leondz/garak ) at 2024-10-28T16:22:59.141907
📜 logging to /home/jemartin/.local/share/garak/garak.log
🦜 loading generator: NIM: microsoft/phi-3-vision-128k-instruct
📜 reporting to /home/jemartin/.local/share/garak/garak_runs/garak.df5ba817-adcb-4660-9c6b-e5507e601dab.report.jsonl
🕵️ queue of probes: encoding.InjectBase64
encoding.InjectBase64 encoding.DecodeMatch: PASS ok on 55/ 55
📜 report closed :) /home/jemartin/.local/share/garak/garak_runs/garak.df5ba817-adcb-4660-9c6b-e5507e601dab.report.jsonl
📜 report html summary being written to /home/jemartin/.local/share/garak/garak_runs/garak.df5ba817-adcb-4660-9c6b-e5507e601dab.report.html
✔️ garak run complete in 7.94s
Comment on lines
+32
to
+34
| DEFAULT_PARAMS = { | ||
| "strict_modality_match": False, | ||
| } |
There was a problem hiding this comment.
This seems reasonable to be in the harness for now, however I see some possible contention as noted in the description questions.
I suspect strict modality matching requirements may lean toward being a responsibility of the probe to define requirements. I suspect the technique employed in the probe will set the requirement for strict match as generators may often support various modalities.
There was a problem hiding this comment.
Interested to hear more
Signed-off-by: Jeffrey Martin <[email protected]>
jmartin-tech
force-pushed
the
generator/vision_nim
branch
from
29f2574
to
a24577a
Compare
Sign up for free
to subscribe to this conversation on GitHub.
Already have an account?
Sign in.
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.
resolves #691
nim.NIMVision, following thehuggingface.LLaVapatternattemptto allow non-string promptsOpenAICompatibleVerification
garak -m nim.NIMVision -n microsoft/phi-3-vision-128k-instruct -p visual_jailbreak.FigStep -g 1 --parallel_attempts 16Questions:
NIMVision._prepare_prompt()?huggingface.LLaVanow needs strict matching. Tagging Enable more complexprompts#602 for good and Prompt Architecture Enhancement for Better Multi-modal Red Teaming #658.