Turn your Claude Code and Codex agent traces into training-ready SFT datasets for small open models

Rasmic · x · 2026-10-16

Ras Cloud's local control plane collects agent sessions from Codex, Claude Code, Cursor etc., using commits/PRs as session boundaries, fixing replaced agent messages and stripping raw logs to produce tokenized training datasets. Screenshot shows an agent-traces set (4,248 train sessions, 13,390 rows, 365.5M tokens, 32.5% assistant) plus 227 eval sessions, and a coding-sft set (390 rows) with breakdowns by harness and host, exportable as JSONL to train small open-weight models.

Original post →

More from coding & agent

coding & agent channel →