feat(data): add When2Call single-turn tool-use dataset preset

Register nvidia/When2Call (train_sft config) in the shared dataset registry.
Since vllm-project#777 that registry is the single source of presets, so the entry serves
both pipelines: off-policy `prepare-data` and the on-policy regeneration
script's `--dataset`.

When2Call is a function-calling SFT dataset (CC-BY-4.0) that also teaches when
NOT to call a tool. Every row is a single user/assistant exchange with no system
turn; the conversation uses the OpenAI `messages` schema (normalized to
`conversations`), and the tool schemas ride along in a parallel `tools` column.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01LRRGNt2T9ktWyFRboNUZNU
Signed-off-by: Ranran Haoran Zhang <ranzhang@redhat.com>

Author: WindChimeRan

docs/cli/response_regeneration.md

@@ -124,6 +124,7 @@ The text presets from the shared dataset registry (`DATASET_CONFIGS` in `specula
 | `magpie`            | `Magpie-Align/Magpie-Llama-3.1-Pro-300K-Filtered` | `train`       |
 | `nemotron`          | `nvidia/Llama-Nemotron-Post-Training-Dataset`     | `chat`        |
 | `open-perfectblend` | `mlabonne/open-perfectblend`                      | `train`       |
+| `when2call`         | `nvidia/When2Call`                                | `train`       |
 
 The registry's multimodal preset, `sharegpt4v_coco`, is **off-policy only** and `--dataset` rejects it. Its turns carry image content parts, which the Chat Completions API rejects, and the pre-tokenized output row has nowhere to keep pixel data. Use it with `prepare-data`.
 

src/speculators/data_generation/configs.py

@@ -49,6 +49,12 @@ def _normalize_nemotron(example: dict) -> dict:
     }
 
 
+def _normalize_when2call(example: dict) -> dict:
+    # When2Call already stores the dialogue as role/content ``messages``;
+    # the parallel ``tools`` column is preserved by the map for tool-aware runs.
+    return {"conversations": example["messages"]}
+
+
 def get_coco_dir():
     return os.getenv("COCO_DIR") or "coco/"
 
@@ -151,4 +157,11 @@ def _normalize_sharegpt4v_coco(example: dict) -> dict:
         hf_path="mlabonne/open-perfectblend",
         split="train",
     ),
+    "when2call": DatasetConfig(
+        name="when2call",
+        hf_path="nvidia/When2Call",
+        subset="train_sft",
+        split="train",
+        normalize_fn=_normalize_when2call,
+    ),
 }