This dataset holds code-search trajectories. The assistant turns were generated by Muse Spark Open (Meta Platforms, Inc.) or by Agents-A1-35B (InternScience, Apache-2.0), and the tool results contain excerpts of source code from public repositories under their own licenses (repo_licenses.csv). By requesting access you agree to the following. (1) Muse Model License, version "Muse Spark Open": "The above copyright notice and this license and the Usage Policy shall be included in all copies or substantial portions of the Software and Derivatives." "Licensee shall prominently display 'Built with Muse Spark Open' in the user interface of each product or service that uses the Software or any Derivatives, or in the accompanying documentation, website, or settings menu." The license defines Derivatives as "any AI model created, trained, modified (including fine-tuned), distilled, or otherwise improved using the Software or any outputs or results thereof", so a model you train on this data is a Derivative and these conditions apply to it. Keep the notice and the license files with any copy of this data. (2) Comply with the Muse Spark Open Usage Policy (Exhibit A of the license, included here), which among other things bars using the model or its outputs to infringe third-party rights. (3) Section 3 of the license (Model as a Service) applies to any model you train on this data and offer to third parties through an API once your aggregate gross revenue with affiliates exceeds US$20 million in any twelve months: you must then notify Meta and sign a separate agreement with Meta. (4) Respect the license of each source repository in repo_licenses.csv: keep its copyright and license notices if you redistribute excerpts. (5) Do not use the data to identify or contact individuals named in the source code. (6) The data is provided as is, without warranty (license section 5). Access is granted by manual approval.

wayfinder-training-data

Built with Muse Spark Open.

These are the 1,372 code-search trajectories used to fine-tune wayfinder-qwen (wayfinder-glimmer was trained on the same examples with a different tokenizer render). Each example is one task: a short request (the subject line of a real commit), the agent's reasoning, its calls to grep, glob, read_file, semantic_search and report, the tool results, and a final ranked list of path:start-end ranges. Access is by request and manual approval.

Size and contents

Item	Count
Examples	1,372
Tasks	342
Repositories	152 (of the 244 in the training split; all 152 are permissively licensed: MIT 83, Apache-2.0 59, BSD-3-Clause 9, BSD-2-Clause 1)
Muse Spark Open trajectories	1,339 (all from self-hosted replicas of the open weights; none from Meta's API or a third-party host)
Agents-A1-35B trajectories	33
Tokens	41.3M with the Qwen3.6 chat template (7.0M assistant tokens, all of which carried loss)
Examples that offered `semantic_search`	1,054 (759 called it at least once)

The counts are of the released files. The original training set had 1,383 examples; 11 were removed before release (8 from one AGPL-3.0 repository and 3 generated through a third-party host). The data is about 155 MB.

Schema

One JSON object per line.

Field	Type	Meaning
`id`	string	`<task_id>\|teacher\|s<seed>` or the Agents-A1 id
`tools_variant`	string	`scout5` (five tools, with `semantic_search`) or `scout4_grep_only`
`messages`	list	Chat messages: one `system`, one `user` (the request), then `assistant` turns with `reasoning_content` and `tool_calls` and `tool` results, ending with a `report` call. Tool-call arguments are objects
`meta.source`	string	`teacher` (Muse Spark Open) or `a1` (Agents-A1-35B)
`meta.task_id`, `meta.repo`, `meta.seed`	string	The task, the `owner/name` repository and the rollout seed
`meta.category`	string	Task category (see below)
`meta.semantic`, `meta.semantic_calls`	bool, int	Whether `semantic_search` was offered, and how many times it was called
`meta.turns`, `meta.n_gold`, `meta.recall`, `meta.all_gold_found`	number	Turns used, number of right files, recall and the all-gold flag from the filter
`meta.tokens`, `meta.assistant_tokens`	int	Example length under the Qwen3.6 template

The tool schemas and system prompt are the ones in the GitHub repository (serve/scout_tool_specs.json, serve/scout_product_prompt.md). Fields removed before release (token ids, loss spans, serving endpoints) are not included. The right-file lists and base commits are not shipped as a separate file.

Task sources and shapes

Training tasks were drawn from public repositories only (commit subjects of public GitHub repositories). Tasks were mined from real commits in public repositories: the commit subject line is the question and the files the commit changed are the answers. The split is by repository, so no repository in this data appears in the development, test or 200-task held-out sets. Seven repositories also appear in an early end-to-end study, whose 17 tasks were excluded from this data. For the 357 training-split tasks used across all versions of this data (the 342 here differ from them by at most 15), the mix is:

Category	Tasks	Language	Tasks	Right files per task	Tasks
implementing_new_features	138	Python	83	1	62
refactoring	74	TypeScript/JavaScript	72	2 to 3	132
paying_tech_debt	54	Rust	71	4 to 6	116
modernize	36	C/C++	41	7 or more	47
cybersecurity_bug_finding	29	Go	38
onboarding	19	Ruby, Other, Swift, C#, Java/Kotlin	15, 14, 9, 8, 6
db_schema_rot, qa_over_codebases	4, 3

The median is 3 right files per task (range 1 to 12), and the request is one line of a median 8 words. Right files include test files. Swift, C# and Java/Kotlin are thin. See the model card for what the model does well on these shapes and where it does not.

License note

This card uses license: other because three sets of terms apply at once: the Muse Model License (version "Muse Spark Open") on the Muse-generated assistant turns, Apache-2.0 on the 33 Agents-A1-35B trajectories, and each source repository's own license on the code excerpts in tool results. The license files travel with the data. Nothing here relicenses a source repository.

How it was built

Muse Spark Open 816B-A42B searched each task with Scout's own explore loop (up to 40 turns, temperature 0.85, top_p 0.95, earlier reasoning kept between turns, the line "Report as soon as the evidence covers the brief; do not exhaust the turn budget." in the system prompt). 5,683 teacher rows went through a filter that kept only runs that found every gold file, opened every file they reported and called report in time, within the task's 75th-percentile turn count and 49,152 tokens, and de-duplicated on task and seed. That left 1,350, of which 1,339 are released. Thirty-three strict all-gold trajectories from Agents-A1-35B were added. Training was LoRA supervised fine-tuning, documented in the model card.

Because of this filter the data is not a sample of what a search agent does. It contains only successful runs, and it favours tasks where the teacher succeeded.

What it must not be used for

The Muse Spark Open license permits using, copying, modifying, fine-tuning, distilling, publishing and distributing the model and its derivatives. Its text, which we quote and do not paraphrase, has no clause that bars using its outputs to train another model, including a competing one. The conditions that do apply are these:

"The above copyright notice and this license and the Usage Policy shall be included in all copies or substantial portions of the Software and Derivatives (defined below). Licensee's use of the Software shall comply with the Usage Policy."
"Licensee shall prominently display 'Built with Muse Spark Open' in the user interface of each product or service that uses the Software or any Derivatives, or in the accompanying documentation, website, or settings menu."
Section 3 (Model as a Service) requires a separate written agreement with Meta before offering Derivatives through an API when the licensee's and affiliates' aggregate gross revenue exceeds US$20 million in any twelve months. It does not apply to internal use, research or any commercial activity that is not Model as a Service.
"Meta claims no ownership of outputs generated by the Software."
The Usage Policy bars, among other things, using Muse Spark Open or its outputs to "engage in or facilitate any action or generate any content that infringes, misappropriates, or otherwise violates any third-party rights", to create malware, and to represent "that the use of Muse Spark Open or outputs are human generated".

Added by us, not by Meta: do not use this data to identify, profile or contact people named in the source excerpts, and do not strip the repository licenses from excerpts you redistribute. Training data from repositories carries each repository's own terms (repo_licenses.csv); this card does not relicense them.

Limitations

One teacher, one filter, one task source. Every example is a run that succeeded, so the data has no recovery from failure beyond what happens inside a successful run.
Task wording is English and is a commit subject line, which is shorter and terser than a user's question. Tasks come from public repositories, and up to 13 tasks come from one repository, so the tasks are not independent samples of the world's repositories.
Language coverage is uneven (see the table above). Swift, C# and Java/Kotlin are thin.
Assistant turns include the teacher's reasoning text, which may contain errors, and tool results contain code excerpts that may contain contributor names, emails or stale secrets from public history. Pre-release scans for secrets, machine paths and internal hosts found only test fixtures and placeholders, and no scan finds everything. E-mail addresses that appear in public source excerpts (author lines, test data) were not redacted.
The one AGPL-3.0 repository in the wider training set, and one repository that could not be license-checked, were excluded.
The category, language and file-count tables are for the 357-task union of training-split tasks used across all versions of this data, not for the 342 tasks released here.

Citation

@misc{wayfinder_training_data_2026,
  title  = {wayfinder training data: code-search trajectories from Muse Spark Open},
  author = {Weller, Henry},
  year   = {2026},
  note   = {Voyage AI. Built with Muse Spark Open.},
  url    = {https://huggingface.co/datasets/hweller1/wayfinder-training-data}
}

Downloads last month: 23

Total file size:

162 MB