Hacker News Reader

Comment by brrrrrm | Hacker News Reader

Curious how this compares to Mistral's release? 3b->7b doesn't seem like a huge leap and Mistral has shown phenomenal results for its size

It's a model 40% of the size of Mistral's designed to be transparent (full training details, datasets & evals here: https://stability.wandb.io/stability-llm/stable-lm/reports/S...) and work on edge devices.

There are improved versions coming but this is the best 3b model and Mistral is the best 7b model.

Thank you for sharing a well-written and transparent training report!

Do you have plans to train/release a fine-tuned 3b chat version or other variants?

how much faster is 3b in practice? Seems like an uncommon size, so it would make sense to have the title of "best 3b" lol

The rule of thumb is that inference speed halves with every doubling of parameter size (and obviously a doubling of memory size).

You can check out real world performances on devices here: https://llm.mlc.ai/

I wonder if it's behind the subscription, but I see no reference source code for StableLM.

> 3b->7b doesn't seem like a huge leap

How so? It's on the general order that seems prevalent (and significant) with LLMs? 3, 7, 15 billion?

Do you have a link/reference about Mistral's performance? I'm interested to read more about it.

looking at the 3b results (here https://github.com/Stability-AI/StableLM#stablelm-alpha-v2 ?), it looks like Mistral (which outperforms Llama-2 13b) is far more powerful

n.b. absolutely no tuning on 3B currently, no RLHF, no instruct, no chat

Does this mean that this is straight-up text completion? That's still pretty useful, you just have to know how to write the prompt, no?