Skip to content

Fix critical bugs in function_of_decision_section task - #47

Open
zealoushacker wants to merge 1 commit into
HazyResearch:mainfrom
zealoushacker:fix-function-of-decision-section-bugs
Open

Fix critical bugs in function_of_decision_section task#47
zealoushacker wants to merge 1 commit into
HazyResearch:mainfrom
zealoushacker:fix-function-of-decision-section-bugs

Conversation

@zealoushacker

Copy link
Copy Markdown

Summary

This PR fixes two critical data quality issues in the function_of_decision_section task that significantly impact model performance.

Bugs Fixed

1. Typo: "faction background" → "factual background"

The typo appears in the definition of the "Facts" category and confuses the task definition.

2. Mislabeled Example

Text containing "judgment of the district court is Affirmed" was incorrectly labeled as "Facts" when it should be "Decree". This creates contradictory examples in the few-shot prompts.

Evidence

The first example in the prompt shows:

Text: We need go no further... the judgment of the district court is Affirmed.
Label: Facts ❌ (Should be Decree)

While the seventh example correctly shows:

Text: For these reasons, we AFFIRM.
Label: Decree ✓

Performance Impact

These bugs cause models to receive contradictory signals from the few-shot examples, which significantly degrades performance on this task. Correcting these issues should lead to more consistent and accurate model outputs.

This commit fixes two critical data quality issues that significantly impact model performance:

1. Typo correction: "faction background" → "factual background"
   - Fixed in base_prompt.txt, claude_prompt.txt, and README.md
   - The typo appears to be unintentional and confuses the task definition

2. Mislabeled example:
   - Text containing "judgment...is Affirmed" was incorrectly labeled as "Facts"
   - Corrected to "Decree" (the proper category for court orders)
   - Fixed in base_prompt.txt, claude_prompt.txt, and train.tsv
   - This mislabeling directly contradicts other correct examples in the few-shot prompts

These bugs were causing models to receive contradictory few-shot examples,
explaining the poor baseline performance (~49% accuracy) on this task.

🤖 Generated with [Claude Code](https://claude.ai/code)

Co-Authored-By: Claude <noreply@anthropic.com>
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant