kilian-group/KBevo-SFT-hotpotqa-6k

KBevo-SFT-hotpotqa-6k

Supervised fine-tuning (SFT) trajectories for the KBevo two-phase policy, generated on HotpotQA. Accompanies [Co-Evolving Structured Knowledge and Reasoning in Language Models](https://arxiv.org/abs/2608.26386) (COLM 2026).

This is the exact SFT dataset used to produce `kilian-group/KBevo-Qwen3-1.7B-SFT` and `kilian-group/KBevo-Qwen3-4B-SFT`, which in turn initialise the `KBevo-Qwen3-1.7B-GRPO` and `KBevo-Qwen3-4B-GRPO` runs.

What's in the file

A single JSON file, trajectories.json, containing one top-level key "examples" whose value is a list of 11,746 SFT records over 5,744 unique HotpotQA questions (~2 rollouts per question). The paper reports this dataset as "6k Gemini two-phase trajectories on HotpotQA" (Appendix B.1).

Each record is a complete two-phase trajectory — question → build a structured KB via inline lookups → derive the answer — expressed as a single target string with KBevo's special DB-lookup tokens (<|db_entity|>, <|db_relationship|>, <|db_return|>, <|db_end|>) inlined.

Record fields

Example (abbreviated annotated_text)

Question:
Ohio State Route 362 runs near the edge of which State Park in Ft. Loramie, Ohio?
Answer:
<thinking> The state park near Ohio State Route 362 in Ft. Loramie, Ohio is
<|db_entity|> Ohio State Route 362 <|db_relationship|> located near state park <|db_return|>...

Provenance

• —Source questions: `hotpotqa/hotpot_qa` (train split, filtered)
• —Generator: Gemini, prompted with the KBevo two-phase templates in `data/prompts/database_creation.json` and `data/prompts/lmlm_agent.json`
• —Post-processing: filtered / selected as described in Section 3 and Appendix B of the paper

Loading

import json
from huggingface_hub import hf_hub_download

path = hf_hub_download(
    "kilian-group/KBevo-SFT-hotpotqa-6k",
    "trajectories.json",
    repo_type="dataset",
)
data = json.load(open(path))
examples = data["examples"]
print(len(examples), "records over", len({r["question"] for r in examples}), "unique questions")
# → 11746 records over 5744 unique questions

# SFT target is the annotated_text field:
print(examples[0]["annotated_text"])

How it was used

Both SFT models were fine-tuned by TRL's SFTTrainer on this file for 3 epochs at effective batch size 48, learning rate 5e-5, cosine schedule with 10% warmup. See the model cards for full hyperparameters and Appendix B.1 of the paper.

Citation

@inproceedings{Noonan2026:co-evolving,
  title         = {Co-Evolving Structured Knowledge and Reasoning in Language Models},
  author        = {Ryan Thomas Noonan and Linxi Zhao and Menghan Xu and Akanksha Sarkar and Mihir Mishra and Dongyoung Go and Kilian Q. Weinberger and Yoav Artzi and Jennifer J. Sun},
  booktitle     = {Proceedings of the Conference on Language Modeling},
  year          = {2026},
  url           = {https://arxiv.org/abs/2608.26386}
}