NexaAIDev

/

Octopus-v2

like 815

Hi, as along as you have a well-defined task. You could formulate the function. You can label the data manually (costly as well), or you can use synthetic data. Please refer to our paper

I read through it and feel like I'm missing information about:

1. How exactly was the vocabulary extended?
   From what I found online, there are multiple ways to extend the vocabulary. So I was wondering what exactly did you guys do?

2. After extending the vocabulary, do the embedding and lm_head layers need to be retrained?
   I'm the paper it's mentioned that after extending the vocabulary, you go through a round of fine-tuning. But from my understanding, fine-tuning won't train the lm_head and embedding layers.
   So what was done to train the above 2 mentioned layers?

Please stay tuned. We will open source code later. For earliest notification, consider to join our waitlist: https://www.nexa4ai.com/contact

Curious to check the opensource codebase out soon to know the details!

Hi @TeddyB

1. We add functional tokens to vocabulary, see
   https://huggingface.co/NexaAIDev/Octopus-v2/blob/main/tokenizer_config.json
2. We will prepare a training pipeline on AWS / Google cloud soon for customized API training requirements

Hi @zackli4ai ,

Thanks for the info, I see the new special tokens added to the tokenizer now

I have some follow-up questions:

1. Have you tried your technique of adding new functional tokens to other base models, like MS Phi-3 Mini or Meta Llama 2 8b?
2. Are you also planning on releasing the dataset you used to train the model?

@TeddyB

1. Yes, Octopus-V4 is based on Phi-3 : https://huggingface.co/NexaAIDev/Octopus-v4
2. We are building a training pipeline on AWS / Google Cloud
   thanks for questions