ScaffoldCoT revamp: 1B-token real coding-agent dataset and a 5-10B model incoming
Saraozte01 · reddit · 2026-10-27
The author of ScaffoldCoT shares updates: tool-calling samples (276,389 examples) were removed after synth-gen harness issues propagated failure modes, and the dataset renamed to Scaffold Chat—now 3,450,159 examples, 13GB, byte-identical otherwise.
- New Scaffold Agent dataset: 1B tokens of real coding-agent traces from open near-frontier models editing real repos in the Pi harness; every tool call actually executed, only the 20% of sessions passing hidden tests, quality grading, two LLM judge passes, and green CI were kept (19,652 records)
- Each record is a self-contained compaction segment (≤64k tokens) with system prompt and tool definitions; trainingsftconfig.yaml documents the setup
- A heavily modified 5-10B model trained with a novel merge approach is in the works
More from coding & agent
- Batch-editing hundreds of drone videos by scripting DaVinci Resolve with Astra — sethlazar · 2026-10-27
- DeepSeek V4.1 Flash lands in Cursor, dominating the cost frontier at $0.60-$0.98 per task — dean_rie · 2026-10-27
- Masking slides for AI: placeholders that survive the model but break the layout — shoumikgoswami · 2026-10-27
- Creator has Claude edit her iPhone Duo review inside Tella's agent video editor — evielync · 2026-10-27
- Ant International and Westlake University release Trajectory2Gym for tool-use agent training — ant-intl · 2026-10-27
- Eight universities propose unified framework for work agents — Zishan Xu · 2026-10-27