docs(regen): dataset choices come from the shared registry

The CLI doc's hand-copied dataset list had already drifted (gsm8k was
never added); point at the shared registry instead of maintaining a
third copy, and document the prepare-data-parity ingestion.

Signed-off-by: Ranran Haoran Zhang <ranzhang@redhat.com>

Author: WindChimeRan

docs/cli/response_regeneration.md

@@ -69,9 +69,9 @@ python scripts/response_regeneration/script.py --dataset magpie
 
 #### Data Arguments
 
-- **`--dataset`** (str, default: `ultrachat`, choices: `magpie`, `ultrachat`) Dataset to process.
+- **`--dataset`** (str, default: `ultrachat`) Dataset preset to process. Accepts any preset from the shared dataset registry (`DATASET_CONFIGS` in `speculators/data_generation/configs.py`); see `--help` for the current list.
 
-- **`--split`** (str, default: dataset-specific) Dataset split. Defaults to `train` for magpie and `train_sft` for ultrachat.
+- **`--split`** (str, default: preset-specific) Dataset split. Defaults to the preset's split.
 
 - **`--limit`** (int, default: `None`) Stop after N rows.
 
@@ -110,10 +110,9 @@ python scripts/response_regeneration/script.py \
 
 ## Supported Datasets
 
-| Dataset   | HuggingFace ID                                    | Prompt Field  | Default Split |
-| --------- | ------------------------------------------------- | ------------- | ------------- |
-| Magpie    | `Magpie-Align/Magpie-Llama-3.1-Pro-300K-Filtered` | `instruction` | `train`       |
-| UltraChat | `HuggingFaceH4/ultrachat_200k`                    | `prompt`      | `train_sft`   |
+Every preset in the shared dataset registry (`DATASET_CONFIGS` in `speculators/data_generation/configs.py`) is supported — the same registry the `prepare-data` pipeline uses, so a dataset defined once works in both. Rows are ingested exactly as in `prepare-data` (`filter_fn`, then `normalize_fn`): system and user turns are kept, and the assistant turns are regenerated. Rows without a usable conversation fall back to the preset's bare-prompt column (`prompt_field`).
+
+Note: `sharegpt4v_coco` additionally requires an image-capable endpoint and local COCO files, matching its `prepare-data` requirements.
 
 ## Output Format