Data Exports and HuggingFace Integration

Jul 8, 2026

Data exports are now live on Traces, along with direct integration with Hugging Face. We believe strongly that our traces are your date. So, we've made it easy to share them, and now, just as easy to understand, move, and reuse that data however you need.

Export everything as a portable archive

From Settings → Exports in any namespace, click New Export to export every trace for offline analysis, fine-tuning, or archival. Exports run as a background job, so the page updates on its own as the status changes. You can optionally narrow an export with a date range and/or a visibility filter before you kick off.

When it's ready, high Download for a .zip name {namespace-slug}-export-{unixtime}.zip. Inside is one normalized JSONL file per entity type, plus a manifest:

manifest.json. Export metadata, schema version, and record counts.
traces.jsonl. One trace per line.
messages.jsonl. One message per line.
parts.jsonl. one part per line (tool calls, thinking blocks, etc.)

Because it's normalized JSONL, it loads directly into DuckDB, pandas, and BigQuery with no preprocessing. Exports stay in your history until you delete them.

Push straight to Hugging Face

Once you've connected Hugging Face from Settings → Integrations, a completed a completed export can be pushed directly to a Hugging Face dataset repo.

New dataset or existing one. In the upload dialog you can pick from a searchable list of your most recently updated datasets (across your personal account and any approved Hugging Face orgs), or create a brand-new dataset repo (public or private).
Optional subdirectory. Drop the files under a path prefix within the repo, or leave it empty to upload to the root.
Single commit, with a link. The JSONL files are pushed as one commit, and a View on Hugging Face link appears with the commit URL when it's done.
Per-trace push, too. You don't have to export the entire namespace. The Download dropdown on any trace has an Export to Hugging Face optoin that uploads the trace as JSONL.

This gives you a direct path from your real trace data into the open-source model community.

Available now

Exports and the Hugging Face integration are available on every personal namespace and organization. On your personal account you can export and push your own data; on an organization, exports and Hugging Face pushes are limited to admins. Head to Settings → Exports at traces.com to get started.

Export everything as a portable archive

When it's ready, high Download for a .zip name {namespace-slug}-export-{unixtime}.zip. Inside is one normalized JSONL file per entity type, plus a manifest:

manifest.json. Export metadata, schema version, and record counts.

traces.jsonl. One trace per line.

messages.jsonl. One message per line.

parts.jsonl. one part per line (tool calls, thinking blocks, etc.)

Because it's normalized JSONL, it loads directly into DuckDB, pandas, and BigQuery with no preprocessing. Exports stay in your history until you delete them.

Push straight to Hugging Face

Once you've connected Hugging Face from Settings → Integrations, a completed a completed export can be pushed directly to a Hugging Face dataset repo.

New dataset or existing one. In the upload dialog you can pick from a searchable list of your most recently updated datasets (across your personal account and any approved Hugging Face orgs), or create a brand-new dataset repo (public or private).

Optional subdirectory. Drop the files under a path prefix within the repo, or leave it empty to upload to the root.

Single commit, with a link. The JSONL files are pushed as one commit, and a View on Hugging Face link appears with the commit URL when it's done.

Per-trace push, too. You don't have to export the entire namespace. The Download dropdown on any trace has an Export to Hugging Face optoin that uploads the trace as JSONL.

This gives you a direct path from your real trace data into the open-source model community.