Hacker News Reader

Since it supports vision, I tested it (an 8 bit quant) against one of my pet problems of coin classification (US coins, but not beauty shots of them, just lots of cell phone images on a piece of paper). It did .. horribly. About 41% accuracy. My local gemma4:27b gets 53% and runs 4x faster. Not too surprising but seemed worth poking at it.

View on HN

The model supports both instruction and RL fine-tuning. It looks like people have started training it already [1]. The RL tuning may be only available to the Cloudflare customers, I couldn't be sure.

1. https://huggingface.co/solanaclawd/clef-solana-research

The biggest upside to using a zero-shot classification model is that it's actually zero-shot, i.e., you don't need to construct task-specific training sets.

If it needs to be fine-tuned, then why not fine-tune smaller and cheaper models that are only as big as the task requires?

I hear what you're saying. But this is not such a model. The article says that various internal teams had to fine-tune Clef to suit their specific needs. They're even rolling out a RL training product line for it.

Fair. Though in this case my actual classifier (dino v3 features fed into a very shallow DNN) works surprisingly well and is insanely faster than using an LLM -- so it really was the one shot performance I was hoping for.

Reply on news.ycombinator.com