buycar/ToolForge-data

ToolForge — Source QA

The multi-hop question answering corpora that ToolForge takes as input, in the exact slices used in the paper.

📄 Paper — arXiv:2512.16149 💻 Code — github.com/Buycar-arb/ToolForge

git clone https://github.com/Buycar-arb/ToolForge.git && cd ToolForge
pip install -e ".[all]"
python download_data.py

Contents

257,901 questions across six corpora, drawn from HotpotQA and 2WikiMultihopQA.

corpus	questions	size	question shape
HotpotQA/bridge_hp	72,991	246.5 MB	bridge: the answer to hop 1 identifies the subject of hop 2
HotpotQA/comparison_hp	17,456	52.5 MB	comparison: "which of X and Y…"
2WikiMultihopQA/compositional_wiki	76,481	134.9 MB	compositional: "the Z of the Y of X"
2WikiMultihopQA/comparison_wiki	51,963	92.4 MB	comparison
2WikiMultihopQA/bridge_comparison_wiki	34,631	70.8 MB	bridge and comparison combined
2WikiMultihopQA/inference_wiki	4,379	9.7 MB	inference over family relations

Record schema

{
  "_id": "5a8b57f25542995d1e6f1371",
  "question": "Were Scott Derrickson and Ed Wood of the same nationality?",
  "answer": "yes",
  "type": "comparison",
  "context": [
    ["Scott Derrickson", ["Scott Derrickson (born 1966) is an American director."]],
    ["Ed Wood", ["Edward Davis Wood Jr. was an American director and screenwriter."]]
  ],
  "supporting_facts": [["Scott Derrickson", 0], ["Ed Wood", 0]]
}

supporting_facts is what makes the pipeline work: it splits context into the passages that contain the answer and everything else. ToolForge treats the first as the evidence a tool call should surface, and runs BM25 over the second to produce realistic distractors — which is how a failed tool call in the generated data can be genuinely, plausibly unhelpful.

Nested columns are stored as JSON strings so they fit in Parquet. Convert to the JSONL the pipeline reads with:

toolforge convert to-jsonl original_data/HotpotQA

Which corpus for which dialogue shape

ToolForge generates 29 dialogue cases in four families, and a corpus can only produce the families its questions are shaped for. Measured by sampling and labelling each corpus:

corpus	→ A	→ B	→ C	→ D	good for
inference_wiki	93%	7%	—	—	A
comparison_wiki	—	100%	—	—	B
comparison_hp	16%	84%	—	—	A, B
compositional_wiki	13%	—	83%	3%	C
bridge_hp	76%	4%	20%	—	A, C
bridge_comparison_wiki	4%	44%	36%	16%	B, C, D

bridge_comparison_wiki is the only corpus that reliably yields D — label roughly six times as much of it to get a comparable number of D samples. Details in docs/choosing-source-data.md.

Citation

@article{chen2025toolforge,
  title={ToolForge: A Data Synthesis Pipeline for Multi-Hop Search without Real-World APIs},
  author={Chen, Hao and Hu, Zhexin and Chai, Jiajun and Yang, Haocheng and He, Hang and Wang, Xiaohan and Lin, Wei and Wang, Luhang and Yin, Guojun and others},
  journal={arXiv preprint arXiv:2512.16149},
  year={2025}
}

Please also cite the original datasets: HotpotQA · 2WikiMultihopQA