stalaei/sdft-tooluse-distil

SDFT Tool-Use — distil format

Pipeline-ready version of the ToolAlpaca tool-use benchmark used in the Self-Distillation Fine-Tuning (SDFT) paper, formatted for direct consumption by the training scripts in distillation_methods.

Splits

Split	Rows	Source
train	4046	SDFT paper's data/tooluse_data/train_data
eval	97	SDFT paper's data/tooluse_data/eval_data

Schema

Column	Type	Description
prompt	list[{role, content}]	Chat-format prompt the student receives. System-role is absent; the tool-use instructions + API docs are in the user message.
teacher_prompt	list[{role, content}]	Chat-format prompt the teacher receives. Contains the golden multi-step trace as an in-context demonstration, followed by "Now answer with a response of your own …".
chat_template_kwargs	dict	Extra kwargs for tokenizer.apply_chat_template. Always {"enable_thinking": false} (Qwen3-compatible).

No answer column. The SDFT-paper recipe is on-policy: the teacher generates fresh from teacher_prompt at each step and that generation becomes the target. For ahead-of-time teacher rollouts, see Self-Distillation/data/swap_answer_column.py.

Load

from datasets import load_dataset

ds = load_dataset("stalaei/sdft-tooluse-distil")
print(ds)
# DatasetDict({
#     train: Dataset({features: ['prompt', 'teacher_prompt', 'chat_template_kwargs'], num_rows: 4046})
#     eval:  Dataset({features: ['prompt', 'teacher_prompt', 'chat_template_kwargs'], num_rows:   97})
# })

Scoring

Use tooluse_regex_scorer from Self-Distillation/evals/scorers.py. It regex-extracts the Action: / Action Input: blocks from the model's response, parses each Action Input as JSON, and scores binary-correct only when both the multiset of action names and the merged argument dict match the golden answer — mirroring the logic in third_party/sdft/eval_tooluse.py from the original SDFT authors.

The reference answer for eval rows is a JSON-serialized list[{"Action": str, "Action_Input": json-string}]. See the golden_answer key in the upstream eval data for the canonical schema.

Reproducing locally

This dataset was built from the SDFT authors' repo (https://github.com/idanshen/Self-Distillation):

# 1) Clone the SDFT authors' repo under third_party/
git clone https://github.com/idanshen/Self-Distillation third_party/sdft

# 2) Build
python Self-Distillation/data/build.py \
    dataset=sdft_tooluse format=distil use_docs=True use_answers=True \
    output=data/builds/sdft_tooluse_train_distil_paper

Data loader: Self-Distillation/data/sdft_tooluse/__init__.py.

Provenance

• Upstream dataset: ToolAlpaca (Tang et al., 2023) — simulated-API multi-step tool-use traces.
• Demonstrations in teacher_prompt: the golden multi-step Thought → Action → Action Input trace from the SDFT authors' train set.
• Evaluation labels: structured golden_answer (list of {Action, Action_Input} records) from the SDFT authors' eval set.

License

Inherits the license of the original SDFT release (https://github.com/idanshen/Self-Distillation) and ToolAlpaca. Please verify upstream before any redistribution or commercial use.