Hacker News Reader

Are they releasing the weights for download? The links to HuggingFace in the readme are giving me 404. This dataset they built on-top of "The Pile" sounds interesting - looking forward to evaluating their claim that 3-7 billion param models can perform on par with 175 billion param GPT-3

View on HN

It looks like it's available here https://huggingface.co/stabilityai/stablelm-base-alpha-7b/tr...

Did they claim this? I didn’t see that claim made in the above post.

"The richness of this dataset gives StableLM surprisingly high performance in conversational and coding tasks, despite its small size of 3 to 7 billion parameters (by comparison, GPT-3 has 175 billion parameters)."

So they did not explicitly say it is comparable, but implicitly compared the two. I'm curious to evaluate what "surprisingly high performance" means exactly.

Reply on news.ycombinator.com