Text-browser, enterprise, office and data-analysis agents
Text/tool agent resources. Notes flag eval-only data, licenses and partial releases. Source audit 2026-10-08; runtime untested.
Viewer • Updated • 1.01k • 9.36k • 136Note Human demonstrations + static benchmark | Train on1,009train tasks; evaluate heldout website/domain splits | License: CC-BY-4.0 data; MIT code; web content rights separate | Use original HTML/text data, not Multimodal-Mind2Web. RawDOM retrieval and candidate grounding needed; static correct action ≠successful live episode. Keep protected testdata out of mixture.
McGill-NLP/WebLINX
Viewer • Updated • 79.8k • 3k • 66Note Human conversational demos + static benchmark | Research-only SFT train; heldout eval | License: CC-BY-NC-SA-4.0 DATA; Apache-2.0 CODE | Mixed screenshot/HTML data but text-only modeling supported; exclude image-required rows. BrowserGym wrapper replays static demonstrations, not necessarily interactive live RL. Code license does not override NC-SA data.
neulab/agent-data-collection
Preview • Updated • 70.7k • 120Note Converted SFT collection + format tooling | SFT adapters, provenance normalization | License: NO umbrella dataset license; per-subfolder licenses: GoBrowseMIT,NNetNavApache2,SynatraCC-BY-SA4 among others | Not independent from upstream datasets. Existing used go-browse-wa,nnetnav-wa,nnetnav-live,synatra. Preserve action/observation boundaries and source licenses; don't duplicate mirrors.
apurvaga/go-browse-wa-raw
Viewer • Updated • 196k • 12.1k • 4Note SFT trajectories or supervised task records | Consider selected train subsets after row-level success and provenance QC | License: MIT declared in ADP subfolder; original HF card metadata missing | Current check establishes metadata and file presence, not task correctness or executable runtime. Mirrors/derived rows are not independent coverage.
stanfordnlp/nnetnav-wa
Viewer • Updated • 50.3k • 249 • 1Note SFT trajectories or supervised task records | Consider selected train subsets after row-level success and provenance QC | License: apache-2.0 HF metadata; source/teacher terms still require review | Current check establishes metadata and file presence, not task correctness or executable runtime. Mirrors/derived rows are not independent coverage.
stanfordnlp/nnetnav-live
Viewer • Updated • 53.7k • 307 • 3Note SFT trajectories or supervised task records | Consider selected train subsets after row-level success and provenance QC | License: apache-2.0 HF metadata; source/teacher terms still require review | Current check establishes metadata and file presence, not task correctness or executable runtime. Mirrors/derived rows are not independent coverage.
oottyy/Synatra
Viewer • Updated • 99.9k • 59 • 4Note SFT trajectories or supervised task records | Consider selected train subsets after row-level success and provenance QC | License: CC-BY-SA-4.0 declared by ADP source subfolder; HF metadata absent | Current check establishes metadata and file presence, not task correctness or executable runtime. Mirrors/derived rows are not independent coverage.
xlangai/AgentTrek
Viewer • Updated • 52.6k • 417 • 24Note SFT web trajectories | Research SFT pending license/text-mode audit | License: UNKNOWN: no HF license metadata; code repo license absent | 52,594dialogue turns ≠52,594episodes. Text accessibility/DOM path must preserve correct element targets; underlying live sites unstable; rights not inferred from publicdata.
jupyter-agent/jupyter-agent-dataset
Viewer • Updated • 95.8k • 19.1k • 182Note SFT executable/simulated notebook trajectories | SFT real-executed E2B subset first | License: Apache-2.0 derived dataset; Kaggle ToS and per-notebook/data license separate | 51,389uniqueexamples rendered thinking/nonthinking twice; not102K independenttasks. executor_type essential: simulatedtool observations are not executionproof. Linked Kaggle filesmay unavailable/private/licenserestricted.
FineEnvs/SmolDataEnvs
RL Environment • Updated • 5.54k • 89.6k • 106Note RL tasks + deterministic grader | Train5000; test250/eval144 heldout | License: MIT derived data/code; underlying Kaggle obligations | 5,394tasks currentcard; heldout intentionally harder. NoLLM in finalgrader. Stochasticdata/libraryversions needpinning; absentKagglefiles aninfrastructureerror notwronganswer. Don't trainheldout.
FineEnvs/SmolDataEnvs-sft
Viewer • Updated • 4.68k • 13.5k • 3Note SFT verified trajectories | SFT warmstart matched to legitimate train taskids | License: MIT derived dataset; Kaggle upstreamrights | 4,677verifiedtrajectories; don'tadd as independentdataalongsideJupyter withoutlineagededup. Tool schema/executor mustmatch inference; open-teacheronly status not proven by generic strongmodels.
FineEnvs/SmolDataEnvs-harbor-train
RL Environment • Updated • 5k • 52k • 19Note RL task dataset packaged for Harbor | Consider selected train subsets after row-level success and provenance QC | License: mit HF metadata; source/teacher terms still require review | 5,000 task family shared with SmolDataEnvs; not new independent data diversity. Separate test/eval bundles stay held out; runtime not validated.
Simia-Agent/Simia-OfficeBench-SFT-30k
Viewer • Updated • 29.6k • 700Note SFT simulated trajectories | Research-only SFT; realenvironment validation needed | License: CC-BY-NC-4.0 per README (HFmetadata absent) | 29,587trainrows. based_on_sample lineage fromOfficeBenchseeds makes evaluationcontamination critical. Simulated APIconsistency !=real application success; code repo https://github.com/microsoft/Simia-Agent-Training.
adyen/DABstep
Viewer • Updated • 1.05M • 156k • 61Note Text benchmark + automaticgrader | Evaluation-only; devtasksforinterface tuningnotfinalscore | License: CC-BY-4.0 dataset; leaderboard/codeartifactterms separately | 450+tasks overCSV/JSON/manuals; noLLMjudge needed. Publicsubmissions/task_scores mustnotbeSFTdata. configtasks splitdefault/dev, notordinarytrain/test corpus.
ServiceNow/WorkArena-Instances
Viewer • Updated • 13 • 29.7k • 16Note EVAL ACCESS ONLY. Gated ServiceNow instance-access bundle, not an SFT/RL task corpus. Hosted instances are restricted to evaluation/research, not training/production. Owner approval required; no terms accepted here. https://github.com/ServiceNow/WorkArena
AmineHA/WebArena-Verified
Viewer • Updated • 1.07k • 1.18k • 2Note EVAL ONLY. Officially linked WebArena-Verified benchmark. A compatibility split named train does NOT make it training data. Keep original tasks held out. https://github.com/ServiceNow/webarena-verified