Posts tagged with

research

I taught a 1.2B model to call tools better. Then my scorer lied to me.
Oct 31, 20265 min read
A LoRA fine-tune of LFM2.5-1.2B on ToolACE: +10 points on exact tool-call arguments, a comparison with 4B and 8B models, and the scoring bug that faked a regression.
Comparing four decision APIs on BANKING77: my first experiment
Oct 8, 20266 min read
My first balanced Verdict experiment: 154 banking requests, 77 intents, and 616 attempts comparing accuracy, failures, latency, and clearly labeled costs.