Modelpublic

rristo/nano-aha-moment-3b

sourceHugging Faceupdated 1y agoView on Hugging Face

0likes7downloads

Model Card

Model Card for rristo/nano-aha-moment-3b

RL model example from here: https://github.com/McGill-NLP/nano-aha-moment

Model Details

Model Description

Implementation of DeepSeek R1-zero style training with:

—Single 80G GPU
—No RL Library
—3B Base Model
—Full Parameter Tuning
—Efficient

Direct Use

For usage see https://huggingface.co/rristo/nano-aha-moment-3b/blob/main/2.0_usage.ipynb example

To train model yourself:

—setup server which has around 120GB of GPU
—you can use this setup: https://huggingface.co/rristo/nano-aha-moment-3b/blob/main/flash_attention_docker.md to have flash attention (dockerfile is available: https://hub.docker.com/repository/docker/rristo/flash-attention-custom/general)
—install necessary dependencies: https://huggingface.co/rristo/nano-aha-moment-3b/blob/main/requirements.txt
—run training script https://huggingface.co/rristo/nano-aha-moment-3b/blob/main/1.0_train_nano_r1_risto.ipynb or https://huggingface.co/rristo/nano-aha-moment-3b/blob/main/1.0_train_nano_r1_risto.py

Training Details

Training Data

https://huggingface.co/datasets/Jiayi-Pan/Countdown-Tasks-3to4