Add IFBench RLVR verifiers environment

[partyplatter08-lab]

Summary

This PR adds a small, self-contained Prime Verifiers environment for IFBench instruction-following RLVR work.

It includes:

• environments/ifbench_rlvr/ with a packaged load_environment() entrypoint for prime eval run / Environment Hub usage
• normalization for both released IFBench eval rows (prompt, instruction_id_list, kwargs) and IF-RLVR training rows from allenai/IF_multi_constraints_upto5 (messages, ground_truth)
• reward functions that reuse IFBench verifier classes, falling back to the allenai/open-instruct IF-RLVR verifier registry for training-only constraint IDs
• README quickstart and environment arguments
• focused tests for dataset normalization, JSONL loading, and fractional verifier reward scoring

Why

The README points users to the released IF-RLVR training data and Open Instruct verifier code, but there is not currently a train/eval-ready Prime verifiers environment in this repo. This gives downstream users a reproducible environment wrapper while keeping the core benchmark code unchanged.

This PR was prepared with AI assistance for the public Prime Intellect Algora IF-RLVR/Bench bounty: https://algora.io/PrimeIntellect-ai/bounties/dderbjHtPwTiGVY4. I have reviewed and tested the changes locally.

Validation

• PYTHONDONTWRITEBYTECODE=1 uv run pytest -> 63 passed, 1 warning
• uv build environments/ifbench_rlvr -> wheel and source distribution built successfully
• PYTHONDONTWRITEBYTECODE=1 uv run --project environments/ifbench_rlvr --python 3.12 python ... -> installed the packaged environment, resolved last_word:last_word_answer through open_instruct.IFEvalG, constructed verifiers.envs.singleturn_env.SingleTurnEnv, and scored a demo response as reward_ok=1.0, reward_bad=0.0
• Dataset registry smoke check on allenai/IF_multi_constraints_upto5[:20] and allenai/IFBench_test[:20] -> missing=[] for both

/claim https://algora.io/PrimeIntellect-ai/bounties/dderbjHtPwTiGVY4

[partyplatter08-lab]

Submitted for the public Prime Intellect Algora IF-RLVR/Bench bounty: https://algora.io/PrimeIntellect-ai/bounties/dderbjHtPwTiGVY4 Current verification on this branch: - PYTHONDONTWRITEBYTECODE=1 uv run pytest -> 63 passed, 1 warning - uv build environments/ifbench_rlvr -> wheel and sdist built successfully The implementation is intentionally scoped as a standalone verifiers environment wrapper around the existing IFBench/Open Instruct verifier logic, so the core benchmark code stays unchanged.

[partyplatter08-lab]

Follow-up verification update: I found and fixed a packaging/runtime issue while doing an end-to-end environment install. The environment now pins open-instruct to a packaged revision that includes open_instruct.IFEvalG, and constrains Python to the dependency-supported >=3.12,<3.13 range.

Additional validation after commit 3d9d406:

• PYTHONDONTWRITEBYTECODE=1 uv run pytest -> 63 passed, 1 warning
• uv build environments/ifbench_rlvr -> wheel and sdist built successfully
• PYTHONDONTWRITEBYTECODE=1 uv run --project environments/ifbench_rlvr --python 3.12 python ... -> installed the packaged environment, resolved last_word:last_word_answer via open_instruct.IFEvalG, constructed verifiers.envs.singleturn_env.SingleTurnEnv, and scored the demo response as reward_ok=1.0, reward_bad=0.0
• Dataset registry smoke check on allenai/IF_multi_constraints_upto5[:20] and allenai/IFBench_test[:20] -> missing=[] for both

This should make the submitted environment reproducible from the package metadata rather than relying on whatever open-instruct HEAD happens to install.

[partyplatter08-lab]

/claim https://algora.io/PrimeIntellect-ai/bounties/dderbjHtPwTiGVY4