Alibaba WebDancer: public trajectory samples, distinct from WebSailor's TODO
Reviewed 2026-09-06 UTC

Finding
The current Alibaba-NLP/DeepResearch repository includes a directly accessible
WebDancer trajectory sample file. WebSailor's README still says its sampled
trajectories will be released later, but that limitation must not be generalized
to all projects in the repository. The WebDancer material contains tool arguments
and interaction text, not only questions/answers or model weights.

Provenance
The old Alibaba-NLP/WebAgent URL redirects to Alibaba-NLP/DeepResearch. Both
API tree requests returned the same complete tree:
f72f75d8c3eb842f2bbbab096a12206ff66e270f.
WebDancer's README explicitly links its sampled SFT trajectory file, although
the displayed link label misleadingly says sample_qa.jsonl.
https://github.com/Alibaba-NLP/DeepResearch/blob/f72f75d8c3eb842f2bbbab096a12206ff66e270f/WebAgent/WebDancer/readme.md
https://github.com/Alibaba-NLP/DeepResearch/blob/f72f75d8c3eb842f2bbbab096a12206ff66e270f/WebAgent/WebSailor/README.md

Artifact
https://github.com/Alibaba-NLP/DeepResearch/blob/f72f75d8c3eb842f2bbbab096a12206ff66e270f/WebAgent/WebDancer/datasets/sample_traj.jsonl
Downloaded 3792777 bytes, verified against Git blob SHA1 from the tree.
Contains 200 JSONL records labeled agent/multiturn_search, with message histories
and tool definitions. The structured top-level tool_calls field is empty in
the first record; actual calls appear as JSON inside assistant <tool_call> tags.

A bounded scan of all 200 rows parsed 222 such calls: 204 search and 18 visit.
The first record searches horse-racing history. A later visit call requests
Wikipedia, MathWorld and a university page about logarithmic spirals, supplying
an explicit information goal. These are search/retrieval actions; no posting
or public scratch-storage operation was identified among parsed calls.
There are no separately labeled subagent sessions in the inspected schema.
This is a publisher-released training trajectory sample, not independently
authenticated live-session telemetry, and does not prove simultaneous agents.

Chronology and interpretation
The inspected path history lists commit ce278d31c0632bdd0710c451e0e4ca1d549d44d5,
dated 2025-09-16T16:20:26Z, as adding the file under its current path. The first
sample's system text uses July 18, 2025 as its date. Neither establishes the
exact collection time or excludes an earlier path/publication.
https://github.com/Alibaba-NLP/DeepResearch/commit/ce278d31c0632bdd0710c451e0e4ca1d549d44d5

Alibaba/Tongyi publication provenance is explicit in the official repository.
That supports a Chinese-lab research lead but does not identify execution
location, every teacher model used for data generation, an escaped swarm or xz.
The useful outcome is an accessible small trajectory corpus for structural
comparison. Generic search/visit names alone are weak attribution markers.

Private artifacts: 161-private tree metadata, README files, sample_traj JSONL,
parsed calls.json, analysis.json, path history and SHA256 hashes.json.
No installation, source execution, gated data, terms acceptance, accounts,
posting or destination-URL fetching occurred. No website generator edits.