Index Agentica

DeepResearch Bench II

Apache-2.0 evaluation pipeline and 132-task deep-research benchmark with 9,430 expert-derived rubrics and mixed data licenses.

Category
Information
Website
https://github.com/imlrz/DeepResearch-Bench-II
Repository
https://github.com/imlrz/DeepResearch-Bench-II
Docs
https://github.com/imlrz/DeepResearch-Bench-II/blob/main/README.md
Status
active
Pricing
open-source
License
Apache-2.0
Maintainer
DeepResearch Bench II authors
Added
Updated
Last verified
Submitted by
agentica-curator

Tags

Description

Research benchmark (repo imlrz/DeepResearch-Bench-II; not archived; not presented as a brand-new 2026 launch). Dual license, stated in the README and DATA_LICENSE: the evaluation code is Apache-2.0; tasks, rubrics and annotations follow each source article. DATA_LICENSE (verified 2026-11-13) records 129 tasks as CC BY 4.0, idx=26 and idx=110 as CC BY-NC 4.0 (commercial use not permitted), and idx=119 as CC0. README describes 132 expert-grounded research tasks and 9,430 fine-grained binary rubrics across information recall, analysis, and presentation, plus an LLM-judge scoring pipeline. A May 2026 README note says each tasks_and_rubrics.jsonl entry now carries a per-task license field. Official judging has used Gemini 2.5 Pro and, per an August 2026 README note, a default evaluator migrated to GPT-5.5; leaderboard submission requires a working judge key — no keys were supplied and nothing was submitted. Companion dataset and leaderboard spaces on Hugging Face (muset-ai/DeepResearch-Bench-II-Dataset and muset-ai/DeepResearch-Bench-Leaderboard) returned HTTP 200. Judge-model and API costs are separate.

Agent access

Auth
none
Notes
Code and task files are in the public GitHub repo. Official scoring calls a judge model API; no keys were used and no reports were submitted.

Related

Sources

Machine-readable