Available sizes

Branch	Bits	Description
8_0	8.0	Maximum quality that ExLlamaV2 can produce, near unquantized performance.
6_5	6.5	Very similar to 8.0, good tradeoff of size vs performance, recommended.
5_0	5.0	Slightly lower quality vs 6.5, but usable
4_25	4.25	GPTQ equivalent bits per weight, slightly higher quality.
3_5	3.5	Lower quality, only use if you have to.

Download instructions

With git:

git clone --single-branch --branch 6_5 https://huggingface.co/MiniLLM_-_VanillaKD-Pretrain-Qwen-500M-exl2 VanillaKD-Pretrain-Qwen-500M-6_5

With huggingface hub:

pip3 install huggingface-hub

To download a specific branch, use the --revision parameter. For example, to download the 6.5 bpw branch: Linux:

huggingface-cli download MiniLLM_-_VanillaKD-Pretrain-Qwen-500M-exl2 --revision 6_5 --local-dir VanillaKD-Pretrain-Qwen-500M-6_5 --local-dir-use-symlinks False

Windows (which apparently doesn't like _ in folders sometimes?):

huggingface-cli download MiniLLM_-_VanillaKD-Pretrain-Qwen-500M-exl2 --revision 6_5 --local-dir VanillaKD-Pretrain-Qwen-500M-6.5 --local-dir-use-symlinks False

Original model description:

library_name: transformers license: apache-2.0 datasets: - monology/pile-uncopyrighted - MiniLLM/pile-tokenized language: - en metrics: - accuracy pipeline_tag: text-generation

VanillaKD-Pretrain-Qwen-500M

paper | code

VanillaKD-Pretrain-Qwen-500M is a 500M model with Qwen achitecture pre-trained with vanilla token-level knowledge distillation on the Pile for 50B tokens. The teacher model is Qwen1.5-1.8B.

We also open-source the tokenized pre-training corpus for reproducibility.

It is used as the baseline for MiniLLM-Qwen-500M

Evaluation

MiniPLM models achieves better performance given the same computation and scales well across model sizes:

Other Baselines

Conventional Pre-Training

Citation

@article{miniplm,
    title={MiniPLM: Knowledge Distillation for Pre-Training Language Models}, 
    author={Yuxian Gu and Hao Zhou and Fandong Meng and Jie Zhou and Minlie Huang},
    journal={arXiv preprint arXiv:2410.17215},
    year={2024}
}