projects
all projects →Denali
activeA completely free AI chat for students — Google OAuth, tool calling, and multi-provider inference. No paywalls, no subscriptions.
research
all research →2025-12-26
BashGemma: Fine-tuning a 270M Parameter Model for Natural Language to Bash Translation
Can a 540MB model — small enough to run on a laptop without internet — reliably translate plain English into bash commands? We fine-tuned Google's smallest open model on 9,000 examples and got from 4.5% to 57.4% accuracy. The trick was in how we trained it, not just what we trained it on.
Jack Large
2025-12-24
Attention Alignment Outperforms Logit Distillation for LLM Compression
When compressing a large language model into a smaller one, should you just teach the small model to match the large model's outputs — or also show it what the large model is paying attention to? We ran the comparison systematically and found that attention maps transfer surprisingly well, cutting error rates by 26%. But combining attention with other internal signals adds nothing.
Jack Large, Madelyn Sarbin