Add IHEval (Instruction Hierarchy Evaluation) benchmark #1463

Summary

Add IHEval (Instruction Hierarchy Evaluation), a benchmark that measures whether a model respects the system > user > tool instruction hierarchy. It is structured as a benchmark group with 9 sub-benchmarks across 4 categories — rule-following, task-execution, safety, tool-use — each evaluated under aligned / conflict / reference settings.

• Upstream: https://github.com/ytyz1307zzh/IHEval (paper: arXiv:2502.08745)
• HF data mirror: https://huggingface.co/datasets/zhihz0535/IHEval

Proposed design

• Benchmark group under nemo_skills/dataset/iheval/ (mirrors the mmau-pro pattern): IS_BENCHMARK_GROUP + 9 sub-benchmarks + a group SCORE_MODULE that emits an overall score, per-category and per-setting averages, and the conflict gap (reference - conflict), per aggregation mode (so iheval:N surfaces pass@1[avg-of-N] at the group level too).
• Data: prepare.py downloads per-variant input_data.json from the zhihz0535/IHEval HF mirror via hf_hub_download; the generated test.jsonl is gitignored (not committed). Rows carry a pre-built OpenAI-style messages list (multi-turn dialogues + tool_calls / tool results), so generation runs with ++prompt_format=openai.
• Scoring: the rule-based scorers are packaged as a standalone, pip-installable Apache-2.0 package — https://github.com/bzantium/iheval — consumed via a thin lazy-import evaluator wrapper, pinned in the Dockerfile. Rationale: upstream IHEval is CC BY-NC-ND with no installable package; the rule-following checkers derive from Google IFEval (Apache-2.0), and the task-execution / safety / tool-use scorers are independent re-implementations of the documented rule-based metrics.
• Metrics: IHEvalMetrics builds on the base pass@k machinery, so iheval:N yields pass@1[avg-of-N], with per-setting and per-variant breakdowns.

Usage

ns eval --benchmarks iheval ...            # all 9 sub-benchmarks
ns eval --benchmarks iheval.safety_hijack  # a single sub-benchmark