stalaei/sdft-science-distil

SDFT Science — distil format

Pipeline-ready version of the SciKnowEval Chemistry Level-3 multiple-choice 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	2674	SDFT paper's data/science_data/train_data
eval	507	SDFT paper's data/science_data/eval_data

Schema

Column	Type	Description
prompt	list[{role, content}]	Chat-format prompt the student receives (system + user).
teacher_prompt	list[{role, content}]	Chat-format prompt the teacher receives. Contains the ground-truth demonstration as an in-context example, followed by "Now answer with a response of your own …". Generating from this produces the on-the-fly distillation target.
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: at training time the teacher generates fresh from teacher_prompt and that generation becomes the per-step 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-science-distil")
print(ds)
# DatasetDict({
#     train: Dataset({features: ['prompt', 'teacher_prompt', 'chat_template_kwargs'], num_rows: 2674})
#     eval:  Dataset({features: ['prompt', 'teacher_prompt', 'chat_template_kwargs'], num_rows:  507})
# })

Scoring

Use science_xml_match_scorer from Self-Distillation/evals/scorers.py. Parses the final <answer>X</answer> XML tag from the model's response and exact-matches against the ground-truth letter (A/B/C/D), mirroring third_party/sdft/eval_science.py:extract_xml_answer from the original SDFT authors.

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_science format=distil use_docs=True use_answers=True \
    output=data/builds/sdft_science_train_distil_paper

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

Provenance

• Upstream dataset: SciKnowEval Chemistry L-3 — MCQ over chemistry knowledge.
• Demonstrations in teacher_prompt: the golden answer + reasoning trace from the SDFT paper's train_data, expert-written by GPT-4o.
• Evaluation labels: the golden letter (A/B/C/D) from eval_data.

License

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