Our team name is Percent_bfd.
This repo contains all the code for the 'NeurIPS Large Language Model Efficiency Challenge: 1 LLM + 1 GPU + 1 Day' competition, including model training reproduction and submission evaluation.
We provide 2 submissions to run eval, both for the A100 tracks.
The folder 'neurips_submission_1' contains the eval docker file for the first submission. And the model weights are uploaded to huggingface Percent-BFD/nips_qwen14b_lora_v9.
The folder 'neurips_submission_2' contains the eval docker file for the second submission. And the model weights are uploaded to huggingface Percent-BFD/nips_qwen14b_lora_v7.
The folder 'docker_train' contains the train docker file to reproduce the model. Additionally, the datasets preparation scripts can be found in its subfolder.
Our training datasets is uploaded to huggingface Percent-BFD/nips_data_v7.
Our training code framework "LLaMA-Effcient-Tuning" originates from hiyouga/LLaMA-Factory. We greatly appreciate him open-sourcing his excellent work.