Fix critical bugs in error_data_expansion.py: tool validation, regex,…

[study8677]

Summary

This PR fixes multiple critical bugs in dataloop/error_data_expansion.py that were affecting data generation quality and system stability.

Key Changes

🛡️ Critical Data Quality Fixes

1. Tool Validation Logic:
   • Fixed validate_generated_sample to actually use the original_tools parameter.
   • Now verifies that generated tool calls exist in the provided tool definitions, preventing the model from hallucinating non-existent tools.
2. Robust Regex Extraction:
   • Updated extract_tools_from_instruction regex to be flexible with whitespace (r'<tools>\s*(.*?)\s*</tools>'). Previously it failed if tags weren't strictly formatted with newlines.
3. Input Parsing Safety:
   • Fixed parse_generated_sample to correctly identify the OUTPUT: delimiter. Previously, if the user input contained the word "OUTPUT:", the content would be wrongly truncated.

🐛 General Bug Fixes

4. Model Name Type Mismatch:
   • Fixed main passing model_name as a string instead of a list, which caused API errors.
5. Message Splitting:
   • Improved convert_from_simple_format to split messages by Role headers instead of double newlines (\n\n), preserving multi-paragraph messages.
6. Safe Dictionary Access:
   • Added safe .get() access for judge_details to prevent crashes when the field is missing.

Impact

These changes ensure that the expanded dataset contains valid, executable tool calls and prevents silent data corruption.

[study8677]

Hi, I found some logic errors in the data expansion script and fixed them in this PR.

The changes cover:

1. Robustness fixes for regex tools extraction and input parsing.
2. Logic fix for the validate_generated_sample function (it was effectively a no-op before).
3. Type fix for the model_name parameter.

Ready for review!