Thomas Carroll

I'm particularly excited about the lower VRAM requirements. Being able to fine-tune on an 11 GB card opens this up to so many more people who don't have access to high-end hardware. The Pokemon example is perfect for demonstrating real-world results too. https://playvio.io/

I found the "structure tax" concept particularly interesting - it makes total sense that smaller models would struggle with the cognitive overhead of maintaining both JSON structure and Python syntax simultaneously. That Mistral-7B example with the malformed string really drives the point home. https://cpsgames.org/