SeaLLMs

/

SeaLLM-13B-Chat

like 60

Follow

SeaLLMs - Language Models for Southeast Asian Languages 86

Thanks ! we just release the SeaLLM-7B-chat model and technical report, will progressively release other models in the next few weeks. Have fun!

Thank you for the 7B version, and I'm eagerly waiting for the 13B version.

When will model 13B be released @nxphi47 ?

What is max length of model ?

@KhanhVu we just released SeaLLM-7B-v2 https://huggingface.co./SeaLLMs/SeaLLM-7B-v2

It has significantly better performance than SeaLLM-13B and surpasses chatgpt on many multilingual reasoning tasks.

Welcome to check it out.

Thanks!

@KhanhVu we just released SeaLLM-7B-v2 https://huggingface.co./SeaLLMs/SeaLLM-7B-v2

It has significantly better performance than SeaLLM-13B and surpasses chatgpt on many multilingual reasoning tasks.

Welcome to check it out.

Thanks!

Thank you for the Team contribution @nxphi47 I see v2 model is based on Mistral and yeah, It's better than v1 based on Llama2

@KhanhVu that’s true, we previously began training seallm-13b before mistral was released.

For v2, we conduct more tuning that enables it to outperform not only mistral-7b-instruct, but also chatgpt-3.5.

@KhanhVu that’s true, we previously began training seallm-13b before mistral was released.

For v2, we conduct more tuning that enables it to outperform not only mistral-7b-instruct, but also chatgpt-3.5.

@nxphi47 Yeah, I think so, because the model is trained on high quality data and carefully prepared (both continual pretraining + SFT + DPO phase). And another question, Was the v2 model trained with the same data compare with v1 model (model underwent carefully designed tuning with reasoning --> any additional reasoning data like math is used to training (SFT or DPO phase) for v2 model ?, the benchmark on several MATH datasets is super amazing)

At the moment, sorry that we don't reveal details about the training process beyond what we stated in the model page yet.