[AhmedTElthakeb]

Summary

This PR brings the C++ IFEval implementation in line with the Python reference by fixing two discrepancies that caused different scoring behavior.

Changes

1. Count characters, not bytes, in tolerance checks

Problem: The ends_with, starts_with, and ConstrainedResponse checks used std::string::size() which counts bytes, while the Python reference uses len() which counts characters. Any non-ASCII character in the tolerance window caused the C++ check to fail where Python would pass.

Example: "My answer is no. ⁇ " is 19 characters (passes Python limit of 16+3) but 21 bytes (fails in C++).

Fix:

• Added utf8_len() to count UTF-8 characters (lead bytes only)
• Added utf8_offset() to get byte offset of the n-th character
• Updated ends_with(), starts_with(), and ConstrainedResponse::verify_() to use character counting

2. Strip whitespace to match Python reference behavior

Problem: The C++ implementation didn't strip whitespace in places where the Python reference does, causing responses with leading/trailing whitespace or blank lines to score differently.

Fix:

• Reordered operations in transform_response() to trim after removing first/last lines (matching Python order)
• Added trim checks in IsFollowed() to prevent blank text from counting as following (strict and loose modes)
• Updated Quotation::verify_() to trim the response before checking for quotes

Files Modified

• common.h: Added UTF-8 helper functions, updated ends_with(), starts_with(), and transform_response()
• types.h: Updated Instruction::IsFollowed(), ConstrainedResponse::verify_(), and Quotation::verify_()

[github-actions]

MLCommons CLA bot All contributors have signed the MLCommons CLA ✍️ ✅