Summary

• add reward_lib.py for scoring prompt/response pairs with the existing IFBench verifiers
• expose a batch make_reward_fn(...) helper for RLVR trainers plus structured RewardResult output for debugging reward shaping
• add run_reward.py as a reproducible local reward smoke runner over prompt/response jsonl files
• document minimal reward-function and smoke-runner examples, with focused tests

Why

The Algora IF-RLVR/Bench bounty calls out a train-oriented integration path for IFBench. The current repository has evaluation scripts, but no small reusable reward function that a training loop can call directly. This keeps the change lightweight by reusing the existing strict/loose verifier implementations and adding a CLI smoke path to prove dataset loading plus reward scoring end to end.

Context: Prime Intellect IF-RLVR/Bench Algora bounty: https://algora.io/PrimeIntellect-ai/bounties/dderbjHtPwTiGVY4

Validation

• uv run pytest -q reward_lib_test.py
• uv run pytest -q
• uv run python -m run_reward --input_data=data/IFBench_test.jsonl --input_response_data=data/sample_output.jsonl --mode=loose --limit=5
• smoke-tested reward_lib.make_reward_fn(...) against data/IFBench_test.jsonl