real error-and-recovery traces for your Tiny Aya SQL agent
hey satish, saw your article "Teaching Tiny Aya to query databases: what fine-tuning fixes, and what it does not". that framing is the right question, and my bet matches yours: a lot of what fine-tuning does not fix is a data problem, not a tuning problem.
I build pretraining data for exactly that gap. 66 recorded sessions of an AI agent fixing real bugs in real GitHub repos. it reads files, writes a fix, runs tests, hits real errors, retries, gets reviewed, and only counts when the tests pass. nothing simulated, and there's a verify script in the repo so you can check every session yourself.
it's built to improve coding, tool calling, agentic execution, and recovery from failure. for a SQL agent the recovery part is the interesting bit: real failed attempts followed by real fixes, which is what querying a database actually feels like
it's for continued pretraining, not SFT. first 100 are free, so all 66 are free today:
https://github.com/Sumama-Jameel/megabit-trajectories
curious what it would do inside your AgentTune pipeline. good luck with the hackathon track.
Sumama