Trajector-SWE-20K contains 20,000 real coding-agent sessions collected with the Trajector plugin (https://trajector.publicai.io) from developers who authorized publication. This repository holds 18,000 of them; a 2,000-session sample is openly downloadable at https://huggingface.co/datasets/publicai-dev/Trajector-SWE-20K-sample. By requesting access you agree to the Trajector terms (https://trajector.publicai.io/terms) and to the access conditions below: No redistribution. You will not re-host, re-upload, mirror or otherwise re-publish the sessions, in whole or in part, on any platform (including other Hugging Face repositories, GitHub, data aggregators or training-data collections); anyone who wants the data must request it here. You may publish models, statistics, annotations and other derived works, provided they do not reproduce the sessions and cite the dataset. You will not attempt to use any credential that may remain in the data, and you will report residual credentials or personal data to opensource@publicai.io.

Trajector-SWE-20K

Collected with the Trajector plugin for coding agents — https://trajector.publicai.io · Paper: Actions Without Reasons (NeurIPS 2026) · This repository holds 18,000 of the 20,000 sessions (access on request). The other 2,000 are openly downloadable at Trajector-SWE-20K-sample.

Trajector-SWE-20K is a dataset of 20,000 real coding-agent sessions, contributed by developers who authorized their sessions for publication, redacted, and released under ODC-BY. Each session is a complete interaction: the system prompt, the tool definitions, and the alternating sequence of user messages, assistant messages, tool calls, and tool results. The sessions span six model generations.

What defines the dataset is what it lacks. Across 774K assistant turns it carries only 17.6 non-empty reasoning blocks per 100 turns, and 37% of sessions contain no non-empty reasoning at all. The actions are recorded; the reasons mostly are not.

Access

The release has 20,000 sessions. This repository holds 18,000 of them and is gated: click Request access above and fill in the short form; requests are reviewed by the maintainers. Approval is per Hugging Face account.
No redistribution. Access is granted under conditions you accept in the request form: the sessions may not be re-hosted, re-uploaded, mirrored or re-published, in whole or in part, on any platform; anyone who wants the data must request it here. Derived works (models, statistics, annotations) may be published if they do not reproduce the sessions and cite the dataset. The ODC-BY licence governs the use of the data; these conditions are the terms on which access to this repository is granted.
The other 2,000 sessions are openly downloadable at Trajector-SWE-20K-sample. They are a hash-ordered draw from the release that went through an additional line-by-line screen (see CHECKS.md). The two parts do not overlap.
The release is drawn from a collection of about 88K consented sessions. The remaining sessions have not completed the same redaction and audit process; access to them is by application to opensource@publicai.io.

Paper

Actions Without Reasons: Filling the Reasoning Gap in a Large-Scale Public Dataset of Real Coding-Agent Trajectories — Qin Wang, Afsoun Owji, Hang Yin, Guangyi Liu. NeurIPS 2026 (link to be added).

Code: https://github.com/PublicAI01/action-grounded-reasoning

At a glance (whole release, 20,000 sessions)


Sessions	20,000 (18,000 here + 2,000 in the open sample)
Assistant turns	774,648
Tool calls	776,815
Estimated tokens	≈0.89B (see note below)
Thinking blocks	168,168, of which 136,531 non-empty (18.8% empty)
Non-empty reasoning per 100 assistant turns	17.6
Sessions with no non-empty reasoning	7,406 (37.0%)
Median session	38 messages, 18 tool calls, ≈30K tokens
90th percentile	188 messages, 94 tool calls, ≈89K tokens

Token counts are estimated from the released (redacted) text as chars/4 + CJK chars/1.5 over the system and messages fields. They are not tokenizer counts, and they are lower than counts on the original transcripts because embedded files (base64 images, archives, and similar) are replaced by placeholders during redaction.

Model composition

Model	Sessions	Est. tokens
claude-opus-4-8	5,858	273M
claude-opus-4-7	5,791	291M
claude-opus-4-6	4,365	147M
claude-sonnet-4-6	2,916	125M
claude-fable-5	634	34M
claude-haiku-4-5	433	16M
other	3	<1M

Where the reasoning is missing

Empty rate of thinking blocks by position of the turn within the session:

Position	Blocks	Empty
Final turn	14,352	16.3%
Second-to-last	11,746	10.8%
3rd to 5th from last	27,526	13.0%
Earlier	114,544	21.3%

The final-turn rate approximates the provider-side omission floor; the excess at earlier positions is consistent with clients stripping reasoning from history on resubmission.

Data format

Sixteen gzip-compressed JSON Lines shards under data/ (18,000 sessions; the open sample has four shards with 2,000). One session per line:

Field	Content
`session_id`	Random UUID assigned at release (original identifiers are not released)
`model`	Model that served the session
`thinking_effort`	Reasoning-effort setting of the session, if any
`system`	System prompt (string or list of content blocks)
`tools`	Tool definitions available to the agent
`messages`	Alternating `user` / `assistant` messages; assistant content is a list of `text`, `thinking`, and `tool_use` blocks, user content carries `tool_result` blocks

There is no train/test split.

How the 20,000 sessions were selected

The consented collection was redacted in full (see Redaction), and every session that the screening flagged was removed. From the sessions that passed:

18,340 sessions belong to the public subset that was defined before redaction (a fixed list of 26,999 sessions; 8,618 of them were removed by screening and 41 were not part of the processed collection).
1,660 sessions were added from the rest of the collection to reach 20,000. They were chosen by sorting candidates by a hash of their identifier and taking the first ones, with no manual selection.

Both parts exclude sessions that touch repositories in SWE-bench Verified, and both keep one session per near-duplicate cluster (clusters are defined by the sequence of tool calls). Duplicates are not labelled in the release because none are included.

Because screening removes whole sessions with suspected residual credentials, inclusion is not independent of content: sessions dense in configuration files, environment variables, or deployment commands are under-represented relative to the collection.

Collection and consent

Sessions come from developers who accepted a data collection and use authorization before contributing. The authorization names public release explicitly among its permitted uses, requires contributors to affirm they are at least 18 (or the local age of majority, if higher), and commits that no contributor identity is published. Every contributor represented here accepted it. The operative clauses are reproduced in the paper's appendix; the full text is at https://trajector.publicai.io/terms

The trajectories were recorded by an agent proxy that users installed and enabled on their own machines; the proxy removed credentials before upload, so the redaction below is the second stage the data went through.

Redaction

Redaction has two parts: deterministic replacement inside sessions, and removal of whole sessions that still look risky afterwards.

Replacement. A frozen, pattern-based pipeline replaces credentials (API keys, tokens, passwords, private keys, connection strings), contact details (email addresses, phone numbers), identity documents, and embedded files. A credential is replaced wherever its value occurs, including restatements in prose, not only where it is assigned. Usernames in home-directory paths receive salted pseudonyms that are consistent across the release, including in file listings and slugged project paths, so directory structure survives while the identity does not. Session identifiers are rewritten to random UUIDs.

The rules were developed over forty audit rounds with an independent reviewer. Each round froze the pipeline (hash timestamped with RFC 3161), drew 1,000 sessions with a seed derived from a public randomness beacon, and had the reviewer read the output. The second part exists because pattern-based replacement alone is not exhaustive.

Whole-session removal. Four screening layers run on the redacted text, and any session flagged by any layer is dropped:

a credential gate (independent regular-expression detectors, distinct from the replacement rules);
an LLM screen: every line that still resembles a credential is sent to a language model with all random-looking values masked by shape descriptors (length, character classes, entropy), so no value leaves the machine; lines judged "credential" or "cannot tell" remove the session;
a cross-session back-scan: every credential value registered during replacement in any session is searched for verbatim in every other session;
rules for password literals in explicit password positions and for dotted high-entropy tokens, added after spot checks of the release.

Across the collection, these layers removed 31% of sessions. The procedures run on the released files, including the extra line-by-line screen of the open sample, are described in CHECKS.md.

Redaction is pattern-based and not exhaustive; residual credentials or personal details may remain, which is one reason this part of the release is gated. Report any to opensource@publicai.io.

A named-entity pass for personal names was tested and rejected: most of its hits carried no personal information and applying it would have damaged the data without removing the risk it was meant to address. Names written in free text are therefore outside what redaction covers, and coverage of them is not measured.

Files

File	Content
`data/part_00..15.jsonl.gz`	18,000 sessions
`MANIFEST.txt`	size and SHA-256 of each shard
`STATS.json`	statistics of the 18,000 sessions in this repository
`STATS_release_20000.json`	statistics of the whole 20,000-session release (the numbers above)
`CHECKS.md`	what was run on the released files

Uses and limitations

Intended for training and analysis of coding agents. It should not be used to identify individuals or infer facts about specific users. Two properties bound that use. Pseudonymization is consistent within the dataset, which supports structural analysis but also leaves per-user trajectories linkable within the release. And redaction of third-party material — names, credentials, or identifiers belonging to people other than the consenting user — works on known formats, so it is not exhaustive.

If you find a residual credential or personal detail, report it to opensource@publicai.io; affected sessions are removed in the next revision.

Withdrawal

A contributor who wants their sessions withdrawn can delete them from the platform's "My data" page or request removal at opensource@publicai.io. Withdrawn sessions are removed from later revisions; copies already distributed cannot be recalled.

Citation

@inproceedings{wang2026actions,
  title     = {Actions Without Reasons: Filling the Reasoning Gap in a Large-Scale Public Dataset of Real Coding-Agent Trajectories},
  author    = {Wang, Qin and Owji, Afsoun and Yin, Hang and Liu, Guangyi},
  booktitle = {NeurIPS 2026 Workshop: Transitioning from Pre-Training to Post-Training},
  year      = {2026}
}

Downloads last month: 283

Total file size:

1.9 GB