Demo Mode: This is a live demo. To submit donations, you need write access to the target dataset repository. Request access or use the manual download option.
Local-first Redaction + Review PR-ready bundles

Transcript Donation Lab

Build a sanitized, reviewable donation bundle without ever uploading raw logs. Everything runs in your browser until you click submit.

Anything merged into the dataset may be copied or mirrored. Removing it later may not fully retract it from downstream use.

Why review your transcripts?

🔍

For yourself

  • Learn from patterns — See which prompts led to better outcomes
  • Spot inefficiencies — Find where you or the agent went in circles
  • Improve your workflow — Understand how you actually use coding agents
  • Audit for leaks — Catch secrets or sensitive data you didn't notice
🤝

For everyone

  • Improve future agents — Real usage data helps train better models
  • Document edge cases — Your tricky bugs help others avoid them
  • Build shared knowledge — Contribute to open research on AI tools
  • Shape the ecosystem — Your data influences how agents evolve
Start small: We recommend uploading just 2-3 transcripts first. Check how they look on HuggingFace, make sure the redaction caught everything, and see if there's anything you didn't think of. Then come back for more.
Work in batches: Your review progress is saved in this browser's localStorage. Come back anytime to continue where you left off.
Note: Progress is stored locally in your browser only. If you clear browser data or switch devices, you'll start fresh. We may add accounts and cloud sync in the future.

Step 0 — Export locally

Run the export script on your machine to create raw_export.zip. The export stays on your computer and is never uploaded until you review and sanitize it.

Download exporter

Python exporter Bash helper PowerShell
python export_transcripts.py --source claude
python export_transcripts.py --source codex
python export_transcripts.py --source opencode

What gets exported?

Claude Code
Codex CLI
OpenCode
New options: Use --dry-run to preview what will be exported, --verbose to see excluded files, or --show-rules to view filtering rules.

Step 1 — Import your exports

Upload one or more export files. Everything is processed locally in your browser.

Claude Code
From export_transcripts.py --source claude
Codex CLI
From export_transcripts.py --source codex
OpenCode
From export_transcripts.py --source opencode

Loaded sources

No files loaded yet.
View file tree
Awaiting import.

Understanding your data

How Claude Code stores data

~/.claude/ contains your session history:

  • history.jsonl — Index of all sessions (prompts, timestamps, project paths)
  • projects/ — Full conversation transcripts as JSONL files

Entry types in session files:

userYour messages to Claude
assistantClaude's responses (text, thinking, tool calls)
systemSystem prompts and context
summaryCompressed context from long conversations
file-history-snapshotSnapshots of file state during edits
How Codex CLI stores data

~/.codex/ contains:

  • history.jsonl — Prompt history (session_id, timestamp, text)
  • sessions/YYYY/MM/DD/ — Full rollout files per session

Entry types in rollout files:

session_metaSession metadata (cwd, git info, model)
response_itemMessages with role (user/assistant) and content
How OpenCode stores data

~/.opencode/ uses SQLite for persistence:

  • *.db / *.sqlite — SQLite database with sessions table
  • Sessions contain messages, metadata, and file changes

Our exporter converts SQLite to JSON for processing.

Step 2 — Sessions

Pick the sessions worth donating. Click any session to view its full content. We only score locally.

Tip: For your first donation, try selecting just 2-3 sessions. You can always come back for more after you've seen how they look on HuggingFace.
How scoring works

Sessions are scored locally to help identify potentially valuable transcripts for donation. Higher scores suggest more useful content.

SignalPointsWhy it matters
Keywords: error, traceback, stack+1.5 eachDebugging conversations
Keywords: diff, patch, git, commit+1.5 eachCode changes
Keywords: test, pytest, npm, yarn, pip+1.5 eachBuild/test tooling
Keywords: tool call, function, stderr, stdout+1.5 eachTool use patterns
Length: 400–8000 chars+2Substantive but focused
Length: >8000 chars-1May be too verbose
Length: <120 chars-1Too short to be useful

Scores are heuristic only. Review content before donating — a low-scored session may still be valuable, and high-scored ones may contain sensitive data.

Step 3 — Configure redaction

Contributor metadata

Defaults to your HuggingFace username if signed in
This preference is recorded in metadata; enforcement depends on downstream users respecting it.

Redaction rules

Select what to automatically redact. Matched content will be replaced with placeholders like [REDACTED_EMAIL].

API keys, AWS credentials, JWT tokens, private keys
sk-abc123... → [REDACTED_API_KEY]
Email, phone, IP addresses, SSN
john@email.com → [REDACTED_EMAIL]
Home directories, usernames in paths
/Users/john/code/ → [REDACTED_PATH]/code/
Replaces entire code blocks with placeholder
Use if code may contain proprietary content
Field selection — Choose which data fields to include

Select which fields to keep in your donation. Essential fields cannot be removed. Path-related fields are excluded by default to protect privacy.

Essential (always included)

Recommended (included by default)

Optional (you choose)

Path/Privacy fields (excluded by default)

Always removed (cannot include)

Live preview of field stripping

See how your field selection affects actual session data.

Session 0 of 0
Before (original) 
 
 
 
 After (fields stripped)   
 
 
 
 
 
 
 
  
 
 
Step 4 — Review & confirm
 
 
 
Redaction summary
 
Run redaction to view summary.
  
 
 
Preview snapshots
 
Sanitized previews will appear here.
 
 
 
 
  
 
  
 
  
 
 
Step 5 — Submit
 
 
 
Submit directly
 
Sign in with HuggingFace to upload your bundle directly to the dataset.
 
 
 
Or download for manual upload
 
Only needed if you want to keep a local copy or prefer to upload via git/web UI.
 
Bundle not built yet.
  
 Manual upload instructions 
 
Option 1: Git workflow
 
 
Option 2: HuggingFace Web UI
 
 
 
 
 
 
 
Sign in to submit
 
 
 
   
 
  Opens Hugging Face login in a new window