datasets
Training and evaluation data, with the modality, task and licence stated up front. Listed live from the Hugging Face Hub.
browsecomp-plus
BrowseComp-Plus
BrowseComp-Plus is a new benchmark for Deep-Research system, isolating the effect of the retriever and the LLM agent to enable fair, transparent comparisons of Deep-Research agents. The benchmark sources challenging, reasoning-intensive queries from OpenAI's BrowseComp. However, instead of searching the live web, BrowseComp-Plus evaluates against a fixed, curated corpus of ~100K web documents from the web. The corpus includes both human-verified evidence documents… See the full description on the dataset page: https://huggingface.co/datasets/Tevatron/browsecomp-plus.browsecomp-plus-corpus
BrowseComp-Plus
Project Page | Paper | Code
BrowseComp-Plus is a new benchmark for Deep-Research system, isolating the effect of the retriever and the LLM agent to enable fair, transparent comparisons of Deep-Research agents. The benchmark sources challenging, reasoning-intensive queries from OpenAI's BrowseComp. However, instead of searching the live web, BrowseComp-Plus evaluates against a fixed, curated corpus of ~100K web documents from the web. The corpus includes both… See the full description on the dataset page: https://huggingface.co/datasets/Tevatron/browsecomp-plus-corpus.browse_compbrowsecomp-plus-trajectoriesBrowseCompLongContext
BrowseComp Long Context
BrowseComp Long Context is a dataset based on BrowseComp to benchmark LLM’s capability to retrieve relevant information from noisy data in its context. It converts the agentic question answering tasks from Browsecomp into long context tasks.
For each of the questions in a subset of BrowseComp, a list of urls are attached. Each url will be paired with an indicator indicating whether the content of the web page is required to answer the question or is… See the full description on the dataset page: https://huggingface.co/datasets/openai/BrowseCompLongContext.browsecompBrowseComp-V3
BrowseComp-V3: A Benchmark Dataset for Multimodal Browsing Agents
A dataset containing 300 samples with encrypted question-answer pairs, images, search trajectories, and sub-goals.
Contents
├── data/
│ ├── train.jsonl # Main dataset (1.44 MB, 300 samples)
│ └── images/ # Referenced images
├── scripts/
│ ├── decryption_script.py # Decrypt entire dataset
│ ├── decrypt_batch.py # Batch decrypt to files
│ ├── encryption_utils.py… See the full description on the dataset page: https://huggingface.co/datasets/Halcyon-Zhang/BrowseComp-V3.BrowseComploom-benchmark-browsecompBrowseComp-ZH
🧭 BrowseComp-ZH: Benchmarking the Web Browsing Ability of Large Language Models in Chinese
BrowseComp-ZH is the first high-difficulty benchmark specifically designed to evaluate the real-world web browsing and reasoning capabilities of large language models (LLMs) in the Chinese information ecosystem. Inspired by BrowseComp (Wei et al., 2025), BrowseComp-ZH targets the unique linguistic, structural, and retrieval challenges of the Chinese web, including fragmented platforms… See the full description on the dataset page: https://huggingface.co/datasets/PALIN2018/BrowseComp-ZH.browsecomp-gpt-oss-120b-260222
browsecomp-gpt-oss-120b-260222
Deep research agent evaluation on data/browsecomp.jsonl (normal split).
Results
Metric
Value
pass@4
46.8%
avg@4
23.9%
Trajectory accuracy
23.9% (1211/5064)
Questions
1266
Trajectories
5064 (4 per question)
Avg tool calls
26.1
Full conversations
✅
Model & Setup
Model
gpt-oss-120b
Judge
gpt-4o
Max tool calls
50
Temperature
0.7
Blocked domains
huggingface.co
Tool Usage… See the full description on the dataset page: https://huggingface.co/datasets/rl-rag/browsecomp-gpt-oss-120b-260222.browsecomp-plus-benchmarkbrowsecomp-no-scroll-gpt-oss-120b
browsecomp-no-scroll-gpt-oss-120b
Deep research agent evaluation on data/browsecomp.jsonl (normal split).
Results
Metric
Value
pass@4
46.0%
avg@4
22.9%
Trajectory accuracy
22.9% (1160/5064)
Questions
1266
Trajectories
5064 (4 per question)
Avg tool calls
27.0
Full conversations
❌
Model & Setup
Model
gpt-oss-120b
Judge
gpt-4o
Max tool calls
50
Temperature
0.7
Blocked domains
huggingface.co
Tool… See the full description on the dataset page: https://huggingface.co/datasets/rl-rag/browsecomp-no-scroll-gpt-oss-120b.browsecomp-trajectoriesbrowsecomp-high-effort-gpt-oss-120b
browsecomp-high-effort-gpt-oss-120b
Deep research agent evaluation on data/browsecomp.jsonl (normal split).
Results
Metric
Value
pass@4
44.1%
avg@4
22.9%
Trajectory accuracy
22.9% (1158/5064)
Questions
1266
Trajectories
5064 (4 per question)
Avg tool calls
55.4
Full conversations
❌
Model & Setup
Model
gpt-oss-120b
Judge
gpt-4o
Max tool calls
100
Temperature
0.7
Blocked domains
huggingface.co
Tool… See the full description on the dataset page: https://huggingface.co/datasets/rl-rag/browsecomp-high-effort-gpt-oss-120b.k-browsecomp
K-BrowseComp
K-BrowseComp is a Korean version of BrowseComp: a web-browsing agent benchmark. Items are grounded in Korean contexts and require retrieving information across multiple Korean websites.
The 300-question verified subset is entirely handcrafted by native Korean speakers and every item underwent thorough manual revision and validation.
📄 Paper: https://arxiv.org/abs/2606.02404
💻 Code: https://github.com/prometheus-eval/K-BrowseComp
Subsets… See the full description on the dataset page: https://huggingface.co/datasets/prometheus-eval/k-browsecomp.Browsecomp-stylebrowsecomp-wrong-ans-exp-filterbrowsecomp-qwen35-35b-a3b-think
browsecomp-qwen35-35b-a3b-think
Deep research agent evaluation on data/browsecomp.jsonl (normal split).
Results
Metric
Value
pass@4
43.0%
avg@4
24.8%
Trajectory accuracy
24.8% (1258/5064)
Questions
1266
Trajectories
5064 (4 per question)
Avg tool calls
41.1
Full conversations
❌
Model & Setup
Model
Qwen3.5-35B-A3B
Judge
gpt-4o
Max tool calls
50
Temperature
0.7
Blocked domains
huggingface.co
Tool… See the full description on the dataset page: https://huggingface.co/datasets/rl-rag/browsecomp-qwen35-35b-a3b-think.browsecomp-plus-selected-tools-analysis-v1
BrowseComp-Plus: Selected Tools Analysis
Side-by-side view of selected tool calls from a reference trajectory alongside the new agent trajectory conditioned on those steps.
Retrieval model: Qwen3-Embedding-8BAgent model: gpt-oss-120bRun: traj_summary_ext_selected_tools_gpt-oss-120b_seed0
Columns
Column
Description
query_id
Query identifier
rationale
GPT rationale for why these k steps were selected from the reference trajectory
selected_indices
Step indices… See the full description on the dataset page: https://huggingface.co/datasets/timchen0618/browsecomp-plus-selected-tools-analysis-v1.browsecomp-high-effort-full-gpt-oss-120b
browsecomp-high-effort-full-gpt-oss-120b
Deep research agent evaluation on data/browsecomp.jsonl (normal split).
Results
Metric
Value
pass@1
20.9%
avg@1
20.9%
Trajectory accuracy
20.9% (264/1266)
Questions
1266
Trajectories
1266 (1 per question)
Avg tool calls
52.9
Full conversations
✅
Model & Setup
Model
gpt-oss-120b
Judge
gpt-4o
Max tool calls
100
Temperature
0.7
Blocked domains
huggingface.co… See the full description on the dataset page: https://huggingface.co/datasets/rl-rag/browsecomp-high-effort-full-gpt-oss-120b.browsecomp-plusBrowseCompPlusRetrieval
BrowseCompPlusRetrieval
An MTEB dataset
Massive Text Embedding Benchmark
BrowseComp-Plus is a reasoning-intensive retrieval benchmark that isolates the retriever component of Deep-Research agents. It provides 830 challenging, multi-hop queries sourced from OpenAI's BrowseComp, evaluated against a fixed corpus of ~100K curated web documents with human-verified evidence labels and hard negatives.
Task category
t2t
Domains
Web
Reference
https://arxiv.org/abs/2508.06600… See the full description on the dataset page: https://huggingface.co/datasets/mteb/BrowseCompPlusRetrieval.browsecomp-oss-env-high-effort-gpt-oss-120b
browsecomp-oss-env-high-effort-gpt-oss-120b
Deep research agent evaluation on data/browsecomp.jsonl (normal split).
Results
Metric
Value
pass@1
19.4%
avg@1
19.4%
Trajectory accuracy
19.4% (245/1266)
Questions
1266
Trajectories
1266 (1 per question)
Avg tool calls
52.5
Full conversations
✅
Model & Setup
Model
gpt-oss-120b
Judge
gpt-4o
Max tool calls
100
Temperature
0.7
Blocked domains
huggingface.co… See the full description on the dataset page: https://huggingface.co/datasets/rl-rag/browsecomp-oss-env-high-effort-gpt-oss-120b.browse_comp_plus_512browsecomp-ctxgraph-30b-rl-fusedent-prod-metrics-v3
browsecomp-ctxgraph-30b-rl-fusedent-prod-metrics-v3
Production run 3/3 FINAL (job vista:826755, 2026-07-13, COMPLETED exit 0, steps 32-40 of 40 — FULL RUN DONE). Vals this job: 0.640(s31 redo, all-time high reading) 0.587(s33) 0.527(s36) 0.593(s39) 0.633(s40 final). Complete 40-step trajectory: baseline 0.527 -> final 0.633, peak readings 0.613/0.640. 10 of 11 vals from s12 onward >= 0.55 (s36=0.527 single dip within +-0.05 val noise). Pre-registered success bar (3 consecutive… See the full description on the dataset page: https://huggingface.co/datasets/lingchensanwen/browsecomp-ctxgraph-30b-rl-fusedent-prod-metrics-v3.browsecomp-ctxgraph-30b-rl-tunedexp-ctxgraph-v1
browsecomp-ctxgraph-30b-rl-tunedexp-ctxgraph-v1
TUNED+EXPLORATION ctxgraph arm, FINAL (runs vista:837826 steps 0-16 + vista:839112 steps 17-33, cancelled at s33 by user decision 2026-07-18 — depth verdict negative). Recipe: tuned reward (de-gated fail shaping, uniqueness x10, anti-farm structural) + deep-groups exploration (rollout n=16, batch=7) + branch_len 16384 + consolidation 3. Full clean-scale val trajectory s0-s33:… See the full description on the dataset page: https://huggingface.co/datasets/lingchensanwen/browsecomp-ctxgraph-30b-rl-tunedexp-ctxgraph-v1.browsecomp-plus-passage-corpus
Passage Corpus for the BrowseComp-Plus Dataset
Total downloads since release, read live from the Hugging Face Hub API (downloadsAllTime)
This repository contains the passage corpus for the BrowseComp-Plus dataset, used in the paper Revisiting Text Ranking in Deep Research, which has been accepted at SIGIR 2026, the 49th International ACM SIGIR Conference on Research and Development in Information Retrieval.
Code: https://github.com/ChuanMeng/text-ranking-in-deep-research
The… See the full description on the dataset page: https://huggingface.co/datasets/grill-lab/browsecomp-plus-passage-corpus.browsecomp-qwen35-35b-a3b-nothink
browsecomp-qwen35-35b-a3b-nothink
Deep research agent evaluation on data/browsecomp.jsonl (normal split).
Results
Metric
Value
pass@4
32.3%
avg@4
16.4%
Trajectory accuracy
16.4% (830/5064)
Questions
1266
Trajectories
5064 (4 per question)
Avg tool calls
36.2
Full conversations
❌
Model & Setup
Model
Qwen3.5-35B-A3B
Judge
gpt-4o
Max tool calls
50
Temperature
0.7
Blocked domains
huggingface.co
Tool… See the full description on the dataset page: https://huggingface.co/datasets/rl-rag/browsecomp-qwen35-35b-a3b-nothink.browsecomp-gptoss-clean-qwen35-sft
BrowseComp GPT-oss SFT Data (Qwen3.5 Format)
Multi-turn SFT training data for Qwen3.5 models, converted from GPT-oss-120B
BrowseComp trajectories. Available in two formats.
Files
OpenAI Messages Format (recommended for general use)
browsecomp-gptoss-clean-full-messages.json — 372 examples, standard messages format with tool_calls
LLaMA-Factory ShareGPT Format
browsecomp-gptoss-clean-full.json — 372 examples, LLaMA-Factory sharegpt format… See the full description on the dataset page: https://huggingface.co/datasets/rl-rag/browsecomp-gptoss-clean-qwen35-sft.