[feature] release of data-juicer model checkpoints in huggingface format

[yxdyc]

Our reference models pre-trained by Data-Juicer are in Megatron format, we provide their download links as follows:

Plz @zhijianma paste the below links later.

• LLaMA-1.3B
  • pre-trained on DJ-refined RedPajama and Pile: 50B token, 100B token, 150B token
  • then fine-tuned on DJ-refined instruct data: DJ-IFT, 4.7B token
• LLaMA-7B, fine-tuned on DJ-refined data (chat, English), DJ-CFT-EN, 40k samples
• LLaMA2-7B-FlagAlpha, fine-tuned on DJ refined data (chat, Chinese), DJ-CFT-ZH, 52k samples

Now, we are converting them into huggingface format, and will upload them in ModelScope and huggingface hub.