Fix PERPLEXITY task#1037
Open
ScottHoang wants to merge 7 commits intohuggingface:mainfrom
Open
Conversation
Collaborator
|
The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update. |
Member
|
thanks for the fix ! happy to review once formatted, don't hesitate to ping me for review :) |
Author
|
Hi @NathanHB, thanks for the response! I have reformatted the file back to its original state. Can you review? 🙏 |
Author
|
hi @NathanHB , not quite sure why it is failing some of the test. I don't think my fix would affect anything beside ppl tasks.... |
Member
|
thanks, taking a look ! |
Member
|
not sure what the issue is, i will take a deeper look this week, if you can, in the meantime run the tests locally ? |
1. Remove KeyError: mc1_targets field only exists in multiple_choice subset, not generation subset used by truthfulqa:gen task 2. Fix backwards answer processing logic that was replacing correct answers with periods instead of preserving answer text These fixes make truthfulqa:gen functional for proper evaluation. Task format: lighteval|truthfulqa:gen|0
Replace exact_match with bleu_1 and bleu_4 metrics for generative TruthfulQA evaluation. Exact match was causing 0 scores because model responses never exactly matched reference strings, even when factually correct. BLEU metrics provide proper n-gram overlap scoring as intended by the original TruthfulQA paper.
… call on metadata dict
- Created simple_truthfulqa_judge.py: Custom task for TruthfulQA generation with GPT-4o judge - Judge evaluates model responses against ground truth answers for truthfulness - Treats "I don't know" responses as correct (truthful uncertainty) - Uses binary scoring: CORRECT (1) or INCORRECT (0) based on factual accuracy - Applied code formatting improvements to llm_as_judge.py 🤖 Generated with [Claude Code](https://claude.com/claude-code) Co-Authored-By: Claude <noreply@anthropic.com>
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.
Fix PERPLEXITY task evaluation crash due to incorrect sampling method detection
Issue
Evaluating perplexity-based tasks (e.g., lighteval|wikitext:103:document_level|0) causes an IndexError crash:
File "lighteval/logging/info_loggers.py", line 298, in aggregate
hash_types: list[str] = list(self.compiled_details.values())[0].hashes.keys()
IndexError: list index out of range
Root Cause
The @cached decorator in cache_management.py incorrectly filters out all PERPLEXITY responses:
- get_sampling_method() cannot distinguish between PERPLEXITY and LOGPROBS responses
- Both have non-empty logprobs, so it returns SamplingMethod.LOGPROBS for PERPLEXITY tasks
- The key difference: PERPLEXITY responses have empty text=[], while LOGPROBS typically has text
- The decorator filters cached results by sampling method
- Looks for SamplingMethod.PERPLEXITY but finds SamplingMethod.LOGPROBS
- All 62 valid responses filtered out → returns empty list []
- DetailsLogger.aggregate() assumes at least one task has details
- Accesses list(self.compiled_details.values())[0] on empty dict
- Raises IndexError
Fixes