datasets
Training and evaluation data, with the modality, task and licence stated up front. Listed live from the Hugging Face Hub.
webchain
WebChain v2
A large-scale, human-annotated dataset of real-world web interaction trajectories for training and evaluating web agents.
[Paper] [Code] [Dataset]
WebChain captures how people complete real tasks on live websites. It is designed for agents that must both identify the correct interface element and reason through a sequence of actions. Each trajectory aligns screenshots, web structure, grounded actions, and reasoning signals instead of treating web navigation as… See the full description on the dataset page: https://huggingface.co/datasets/webagentlab/webchain.webchain-legacy
WebChain
WebChain is a large-scale, human-annotated dataset of real-world web interaction trajectories for training and evaluating GUI agents and web agents. WebChain contains 31,725 trajectories, 317,993 steps, and 428 unique domains. Its core contribution is a Triple Alignment of visual context, structural context, and action grounding, enabling supervision for both spatial grounding and long-horizon planning.
Paper: https://arxiv.org/abs/2603.05295
Open access… See the full description on the dataset page: https://huggingface.co/datasets/webagentlab/webchain-legacy.AFM-WebAgent-RL-Dataset
Data Introduction
This dataset serves as the core training data for Agent Foundation Models (AFMs), specifically designed to elicit end-to-end multi-agent reasoning capabilities in large language models. Built on the novel "Chain-of-Agents (CoA)" paradigm, the dataset leverages a multi-agent distillation framework to transform collaboration processes from state-of-the-art multi-agent systems into trajectory data suitable for supervised fine-tuning (SFT), simulating dynamic… See the full description on the dataset page: https://huggingface.co/datasets/PersonalAILab/AFM-WebAgent-RL-Dataset.AFM-WebAgent-SFT-Dataset
Data Introduction
This dataset serves as the core training data for Agent Foundation Models (AFMs), specifically designed to elicit end-to-end multi-agent reasoning capabilities in large language models. Built on the novel "Chain-of-Agents (CoA)" paradigm, the dataset leverages a multi-agent distillation framework to transform collaboration processes from state-of-the-art multi-agent systems into trajectory data suitable for supervised fine-tuning (SFT), simulating dynamic… See the full description on the dataset page: https://huggingface.co/datasets/PersonalAILab/AFM-WebAgent-SFT-Dataset.agent-web-index
Agent Web Index — how much of the web can AI assistants actually read?
48,154 domains measured live. 25% of them cannot be read by at least one of
ChatGPT, Claude, Perplexity or Gemini. Updated daily. Live index: https://shop.lumnika.com/ai-readiness/
Every row here is the result of real HTTP requests, not an estimate and not a re-publication of
someone else's crawl: each domain's homepage is requested once as a browser and once as each of the
published AI crawler user-agents… See the full description on the dataset page: https://huggingface.co/datasets/DeusHorizon/agent-web-index.st-webagentbench
A Benchmark for Evaluating Safety & Trustworthiness in Web Agents
Accepted at ICLR 2026
Overview
ST-WebAgentBench is a policy-enriched evaluation suite for web agents, built on BrowserGym. It measures not only whether agents complete tasks, but whether they do so while respecting safety and trustworthiness (ST) policies — the constraints that govern real enterprise deployments.
The… See the full description on the dataset page: https://huggingface.co/datasets/ST-WebAgentBench/st-webagentbench.agentic-web-cheatsheets
Bowmark: Agentic Web Cheatsheets — Free Sample
Bowmark indexes how websites actually work, for AI agents.
Each row is a cheatsheet for one task on one site: the behavioral gotchas you
only learn by driving the site, a deep-link shortcut where one exists, and a
verification stamp saying how many times it worked and as of when. Every row
was run end-to-end and proven to work — that's the gate to be included.
This repository is a free, curated sample — the strongest… See the full description on the dataset page: https://huggingface.co/datasets/bowmark-ai/agentic-web-cheatsheets.webagent-dom-with-imagesweb-search-agent-sft-traces
Dataset Card for "web-search-agent-sft-traces"
More Information needed
lumos_web_agent_plan_iterative
🪄 Agent Lumos: Unified and Modular Training for Open-Source Language Agents
🌐[Website]
📝[Paper]
🤗[Data]
🤗[Model]
🤗[Demo]
We introduce 🪄Lumos, Language Agents with Unified Formats, Modular Design, and Open-Source LLMs. Lumos unifies a suite of complex interactive tasks and achieves competitive performance with GPT-4/3.5-based and larger open-source agents.
Lumos has following features:
🧩 Modular Architecture:
🧩 Lumos consists of planning, grounding… See the full description on the dataset page: https://huggingface.co/datasets/ai2lumos/lumos_web_agent_plan_iterative.lumos_web_agent_ground_iterative
🪄 Agent Lumos: Unified and Modular Training for Open-Source Language Agents
🌐[Website]
📝[Paper]
🤗[Data]
🤗[Model]
🤗[Demo]
We introduce 🪄Lumos, Language Agents with Unified Formats, Modular Design, and Open-Source LLMs. Lumos unifies a suite of complex interactive tasks and achieves competitive performance with GPT-4/3.5-based and larger open-source agents.
Lumos has following features:
🧩 Modular Architecture:
🧩 Lumos consists of planning, grounding… See the full description on the dataset page: https://huggingface.co/datasets/ai2lumos/lumos_web_agent_ground_iterative.web-search-agent-sft-traces2
Dataset Card for "web-search-agent-sft-traces2"
More Information needed
embodied_web_agent_outdoor_trajectorymini_rm_benchmark_for_web_agent
Dataset Card for "mini_rm_benchmark_for_web_agent"
More Information needed
webagent_policy_rationale_formattedqwen-web-agentwebagent-r1-distillweb-agent-trajectory-testweb-agent-trajectory-multimodal-testembodied-web-agent-geoguessrwebagent-domoumi-web-agentkoen-web-agent-sft-mixes
Korean-English Web Agent SFT Mixes
브라우저 GUI 에이전트 SFT 용 한국어·영어 혼합 데이터. 언어 비율만 다르고 나머지는
동일하게 통제된 4개 구성이라, 비율이 성능에 미치는 영향을 직접 비교할 수 있다.
구성
config
ko : en
스텝
궤적
ko_only_20k
10 : 0
20,001
3,042
mix_ko_en_5050
5 : 5
20,008
2,663
mix_ko_en_2080
2 : 8
20,003
2,436
en_only_20k
0 : 10
20,009
2,274
비율은 궤적 수가 아니라 스텝 수 기준이다. 스텝 하나가 학습 샘플 하나인데
한국어 궤적은 평균 6.6스텝, 영어는 8.7스텝이라, 궤적 수로 5:5 를 맞추면 실제
gradient 기여가 5:5 가 되지 않는다. 궤적은 절대 쪼개지 않는다.
네 구성의 표본은 서로 중첩된다.… See the full description on the dataset page: https://huggingface.co/datasets/GUIAgentt/koen-web-agent-sft-mixes.web-agent-mind2webweb-agent-titlesweb-agent-multitask-flat-failed-stepsweb-agent-multitask-runs-failedwebagent_policy_formattedweb-agent-multitask-flat-successful-shortest-stepsagentic_ii_agent_Qwen3_coder_prompt_web_bench