[Dataset] [Tool call] Add When2Call single-turn tool-use dataset

[WindChimeRan]

Purpose

Add the When2Call dataset (nvidia/When2Call, train_sft config) as a predefined preset, to support tool-use / function-calling speculator training.

why:

• We're adding on-policy tool-call regeneration; a single-turn tool-use dataset is the natural training fixture for it.
• When2Call is a single-turn function-calling SFT dataset (originally for post-training) that also teaches when not to call a tool — useful signal for a draft model.
• License CC-BY-4.0 (permissive). Verified no overlap with the nvidia/SPEED-Bench eval set — SPEED-Bench contains no function-calling data, and When2Call is not among its 24 source datasets.

Rebased onto main after #777, the followup that collapsed the two dataset registries into one. This is now a single entry in src/speculators/data_generation/configs.py — since #777 the regeneration script consumes that same registry, so the preset reaches both pipelines (prepare-data and --dataset) by construction, with no change to scripts/response_regeneration/script.py. docs/cli/response_regeneration.md gains the matching row.

Warning

Draft: on-policy regeneration is the wrong signal for this preset.

Because #777 makes every registry preset a valid --dataset by construction, this entry also enables --dataset when2call. That is not usable yet: When2Call rows are exactly [user, assistant] with no system turn, and the tool schemas live in a parallel tools column that never enters the conversation. On-policy regeneration would therefore send the model a bare user prompt with no tools attached and regenerate a prose answer — precisely the wrong supervision for a dataset whose purpose is teaching when not to call a tool. Off-policy prepare-data is unaffected.

Holding this as draft until the tool-aware regeneration path (#750) lands, or the preset is gated off-policy-only the way #777 gates sharegpt4v_coco.

Notes for reviewers

• configs.py: _normalize_when2call maps the OpenAI messages column to conversations. Once feat(data): declarative schema normalization for prepare-data (RFC #583) #688 (automatic messages→conversations rename) lands, this normalizer becomes redundant and the entry can be simplified to a bare DatasetConfig. Kept for now so the preset works on current main.
• The tools column (a list of JSON-encoded schemas) is preserved as-is; it is not consumed by the offline path on main yet — it is there for the tool-aware regeneration path. [Regen][tool call] Support tool calls in on-policy response regeneration #750

Tests

• make quality (ruff check, ruff format, mdformat, mypy): clean.
• pytest tests/unit/scripts tests/unit/data_generation: 49 passed.
• Schema verified against the real data file — 2485 rows parsed from train/when2call_train_sft.jsonl: every row is exactly [user, assistant] (role/content) alongside a parallel tools column. No multi-turn rows, no system turns. Config/split resolve confirmed against the Hub API (train_sft config → train split, CC-BY-4.0).
• Registry sanity: _normalize_when2call maps a real When2Call row to conversations while the tools column survives the map; prepare_row() on that row yields [user]; when2call appears in the --dataset CLI choices with script.py byte-identical to main.

Checklist

I have filled in:

• The purpose of the PR, such as "Fix some issue (link existing issues this PR will resolve)".
• The test plan/results, such as providing test command and pasting the results.
• (Optional) The necessary documentation update.
• I (a human) have written or reviewed the code in this pr to the best of my ability.

[coderabbitai]

📝 Walkthrough

Walkthrough

Changes

When2Call dataset support

Layer / File(s)	Summary
When2Call normalization and registry src/speculators/data_generation/configs.py	Adds When2Call normalization from messages to conversations and registers nvidia/When2Call with the train_sft subset and train split.
Regeneration CLI integration scripts/response_regeneration/script.py	Adds when2call as a supported dataset using the existing regeneration pipeline configuration.

🚥 Pre-merge checks | ✅ 5

✅ Passed checks (5 passed)

Check name	Status	Explanation
Docstring Coverage	✅ Passed	No functions found in the changed files to evaluate docstring coverage. Skipping docstring coverage check.
Linked Issues check	✅ Passed	Check skipped because no linked issues were found for this pull request.
Out of Scope Changes check	✅ Passed	Check skipped because no linked issues were found for this pull request.
Title check	✅ Passed	The title clearly summarizes the main change: adding the When2Call single-turn tool-use dataset preset.
Description check	✅ Passed	The description is directly related to the changes and explains the new When2Call preset and its intended use.

✨ Finishing Touches

🧪 Generate unit tests (beta)

• Create PR with unit tests

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands.}

[orestis-z]

LGTM. Clean, well-scoped addition that follows the existing dataset preset patterns. Config matches the actual HuggingFace dataset schema (train_sft subset, train split, messages + tools columns). Recommend approving.

🤖 Generated with Claude Code using the /pr-review skill

[mergify]

This pull request has merge conflicts that must be resolved before it can be
merged. Please rebase the PR, @WindChimeRan.

https://docs.github.com/en/pull-requests/collaborating-with-pull-requests/working-with-forks/syncing-a-fork

[WindChimeRan]

close as low value to speculator training.