I’ve been experimenting with next-edit prediction for a while and wrote up how we trained the edit model that powers our Tab completion feature. This post is part of a broader series where we share how we built this feature from the low-level modeling right up to the editor extension.
The cool part is we fine-tuned Gemini Flash Lite with LoRA instead of an OSS model, helping us avoid all the infra overhead and giving us faster responses with lower compute cost.
1 comment
[ 2.6 ms ] story [ 9.9 ms ] threadThe cool part is we fine-tuned Gemini Flash Lite with LoRA instead of an OSS model, helping us avoid all the infra overhead and giving us faster responses with lower compute cost.