ifeval_ood: read the public IFBench_test dataset

[abhishekraok]

ifeval_ood reads allenai/IFBench_test2, which is private and gated behind the allenai org's fine-grained-token policy. Anyone holding a classic HF token gets a load failure, so the task is unrunnable — this is what has IFBench stuck on the migration tracker.

allenai/IFBench_test is the benchmark's canonical published set, is public and ungated, and carries the same data. Switching to it makes the task runnable without changing any score.

Verification that the swap is score-neutral

IFBench_test2 can't be read from an unprivileged environment, so its contents were recovered from the request file of an oe-eval run that could read it (result dataset 01M0X9N1R1VDM1YEH9J16NTFGX, experiment 01M0X8SCW6QS5V5BSXP73RKP1V) and diffed against the public set on key, prompt, instruction_id_list and non-null kwargs:

public  allenai/IFBench_test : n=300
recovered IFBench_test2      : n=300
row-order mismatches         : 0
content hashes               : equal

Task level, after the switch: 300 instances load, all 58 distinct instruction types resolve in the ifbench registry, and all 300 score end to end with zero verifier build errors.

The same identity was confirmed independently against ifbench/data/IFBench_test.jsonl in the official allenai/IFBench repo.

Notes

• oe-eval hardcodes IFBench_test2 too (oe_eval/tasks/oe_eval_tasks/ood_ifeval.py:37). Cross-harness IFBench parity wants the same switch there — separate repo, separate change.
• This does not unblock the whole ifbench suite. The two ifeval_mt tasks read VGraf/ifeval_mt, which was already accessible.
• Two docstrings referring to the private copy are updated alongside the data source.
• Worth someone at Ai2 establishing why the private duplicate exists before it's relied on again.

Full suite 1957 passed, ruff/ty clean.

🤖 Generated with Claude Code

https://claude.ai/code/session_01LMFjgUqHjQA3aCeB4QmEHZ

[undfined]

lgtm