docs: fix dataset_name override path in grpo task mapping note

[yuki-97]

Summary

• Fix one outdated config path reference in the GRPO task–dataset mapping note: config.data.dataset_name → config.data.train.dataset_name. Training configs nest dataset_name under data.train (see examples/configs/grpo_math_8B.yaml, sft_openmathinstruct2.yaml), so the previous wording would silently mislead anyone using it as a Hydra override.

Closes #2719.

Note on eval docs

Eval configs (examples/configs/evals/*.yaml) use a flat data.dataset_name key with no train sub-key — see eval.yaml, gpqa_eval.yaml, math_eval.yaml, mmau.yaml, etc. So all eval-context references in the docs (docs/guides/eval.md, docs/about/evaluation.md, the eval section of docs/guides/grpo-audio.md, and the run_eval.py example in docs/guides/sft-openmathinstruct2.md) are correct as-is and intentionally left unchanged.

Of the 11 data.dataset_name= occurrences in docs/, 10 are eval (correct) and only the one in docs/guides/grpo.md describes training (incorrect, now fixed).

Test plan

• Docs-only change; no runtime impact.

[copy-pr-bot]

This pull request requires additional validation before any workflows can run on NVIDIA's runners.

Pull request vetters can view their responsibilities here.

Contributors can view more details about this message here.

[yuki-97]

/ok to test 40f39af