Ouro multi-source session dataset

The private Kaggle dataset lanternfounder/ouro-claude-sessions must contain training-data.claude-combined.json. Kaggle mounts it at:

/kaggle/input/ouro-claude-sessions/training-data.claude-combined.json

The historical name is retained for compatibility. The JSON array now combines:

the existing scrubbed Claude training corpus;
successful Claude Code tool trajectories;
successful Codex function calls with matching function-call outputs;
ChatGPT export turns with authentic tool calls and matching tool results.

Plain ChatGPT conversations are not labeled as tool examples. Missing or errored tool results are rejected. Every generated row includes source, hashed session, tool-use, and hashed call provenance under metadata.

Build locally

Generated session data is private and gitignored:


python scripts/build_claude_session_dataset.py `

  --base models/lantern-sigma0-coder/training-data-rebalanced.jsonl `

  --chatgpt-source C:\private\chatgpt-export\conversations.json `

  --out models/lantern-sigma0-coder/training-data.claude-combined.json

The builder discovers Claude under CLAUDE_PROJECTS_DIR (or ~/.claude/projects) and Codex under CODEX_HOME (or ~/.codex). ChatGPT has no implicit filesystem scan: pass --chatgpt-source or set CHATGPT_SESSION_EXPORT.

Automatic Claude and Codex discovery is restricted to session paths/working directories containing lantern-os. Change that marker with --match; explicit source paths remain available for deliberately scoped imports.

Run with --dry-run first. The summary reports discovered files, source rows, tool rows, skipped/error records, redactions, and deterministic deduplication. Review the summary and private output before uploading it to Kaggle.

No raw transcript or generated combined dataset belongs in Git.