This repository’s own code is licensed under the MIT License (see LICENSE).
It also redistributes a small amount of third-party data, listed below with
its original license and attribution.
eval/questions_ext.json contains 24 research questions (ids e01–e19,
e21–e25; e20 was intentionally skipped to avoid reusing the id of a
removed question — see below) borrowed verbatim from the DeepResearch
Bench project. Of these, e11–e19 and e21–e25 (14 questions) are
used in this repo’s main n=24 comparison (added to the 10 self-authored
questions in eval/questions.json); e01–e10 were used in earlier
exploratory generalization checks documented in EXPERIMENTS.md and are
not part of the main comparison table. (A 20th borrowed question,
box-office rankings, was tried and removed — see EXPERIMENTS.md for why.)
eval/questions_ext.json carries a "source"
field naming the exact upstream question id it was copied from
(e.g. "DeepResearch-Bench#6 (Apache-2.0, Ayanami0730/deep_research_bench)").No other content from DeepResearch Bench (code, scoring rubric, or other question sets) is included in this repository — only the question text itself, used here as a held-out evaluation set for a different purpose (comparing citation-faithfulness/report-quality tradeoffs across pipeline designs on a weaker base model) than the original benchmark’s intended use.
Per Apache-2.0 §4, this NOTICE file preserves attribution for the redistributed
content; it does not itself relicense this repository’s own code, which
remains MIT (see LICENSE).