HackerTrans
TopNewTrendsCommentsPastAskShowJobs

mluo

no profile record

Submissions

DeepScaleR: Surpassing O1-Preview with a 1.5B Model by Scaling RL

pretty-radio-b75.notion.site
19 points·by mluo·last year·0 comments

comments

mluo
·last year·discuss
Check out one of my prior work: https://stylus-diffusion.github.io/

This work scales up selection/routing over many models/LoRAs
mluo
·last year·discuss
For quantization, very big impact for small models, can drop at much as 10% on AIME. Our model does best on bfloat16 ;)

Come checkout our repo at: https://github.com/agentica-project/deepscaler
mluo
·last year·discuss
It's simply bc the model is small (1.5B), making it sensitive to weight perturbations
mluo
·last year·discuss
Think there are some people who made GGUFs as branches of our model, try it out!

https://huggingface.co/models?other=base_model:quantized:age...
mluo
·last year·discuss
Nice, very glad to see it works! Small models are very sensitive to the dtype :(
mluo
·last year·discuss
Try bfloat16! We have a bug where the model was saved as fp32.
mluo
·last year·discuss
We beat O1-preview and even many other 7B models over many math benchmarks, which was TEST set (not in training set at all).

If you want to make the model fully generalist, feel free to train it over coding datasets (such as RL with passing unit tests as reward).
mluo
·last year·discuss
One of the authors here....

This is not a Chinese model, btw I'm American
mluo
·last year·discuss
Hi, one of the lead authors for this work.

We recommend using Bfloat16 (not fp16), quantization for small models can really hurt performance!
mluo
·3 years ago·discuss
Alpaca Llama Vicuna -> Gorilla

Chad move
mluo
·3 years ago·discuss
With inflation in mind, wouldn't there a larger gap between Ray's sort and the previous WR holder?