The most useful fine-tuning data is real traces: sessions, agent runs, tool calls
ziv_ravid · x · 2026-11-20
Ziv Ravid's thread (posts 5-6): the most useful data is traces of real activity — user sessions, agent runs, tool calls. They don't need to be labeled, or labels can be sparse with no answer for each step. Keep a small dataset aside to verify you didn't break anything.
Earlier advice: start with a few hundred real production tasks, look at the data manually to build intuition about what good outcomes look like, then define metrics — that becomes much easier afterward.
More from coding & agent
- QuixiInfer author: tell Claude 'defeat is not an option' to make it write kernels for your GPU — QuixiAI · 2026-11-20
- NormGym benchmark shows frontier agents can't infer unspoken community norms — xuanalogue · 2026-11-20
- Dev argues terminal/tmux is the wrong form factor for managing 5-10 AI agents — DuaneJRich · 2026-11-20
- Volcano launches Agent Cloud: self-healing deploy platform built for coding agents — rohanpaul_ai · 2026-11-20
- Dusk bot: open-source Grok Bot alternative reuses your ChatGPT/Claude/Grok subscriptions, 476 GitHub stars — EXM7777 · 2026-11-20
- Theo's Restaurant Analogy Nails Agentic UX: The Whole Kitchen Waits on Your Fries Decision — seatedro · 2026-11-20