DeepScaleR: Surpassing O1-Preview with a 1.5B Model by Scaling RLpretty-radio-b75.notion.site19 points·by mluo·l’année dernière·0 comments
mluo·l’année dernière·discussCheck out one of my prior work: https://stylus-diffusion.github.io/This work scales up selection/routing over many models/LoRAs
mluo·l’année dernière·discussFor 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·l’année dernière·discussIt's simply bc the model is small (1.5B), making it sensitive to weight perturbations
mluo·l’année dernière·discussThink 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·l’année dernière·discussNice, very glad to see it works! Small models are very sensitive to the dtype :(
mluo·l’année dernière·discussWe 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·l’année dernière·discussOne of the authors here....This is not a Chinese model, btw I'm American
mluo·l’année dernière·discussHi, one of the lead authors for this work.We recommend using Bfloat16 (not fp16), quantization for small models can really hurt performance!
mluo·il y a 3 ans·discussWith inflation in mind, wouldn't there a larger gap between Ray's sort and the previous WR holder?
This work scales up selection/routing over many models/LoRAs