#Any downsides in fine-turning models with non-proprietary chat templates?

5 messages · Page 1 of 1 (latest)

robust prism
#

Are there any downsides in fine-turning models with non-proprietary chat templates?

As en example, if I fine-tune llama3 with the Alpaca chat template, will it be performing worse than using the original llama3 chat template?

For reference, llama3 uses:


You are a helpful AI assistant for travel tips and recommendations<|eot_id|><|start_header_id|>user<|end_header_id|>

What can you help me with?<|eot_id|><|start_header_id|>assistant<|end_header_id|>```

While the Alpaca one is:

```Below is an instruction that describes a task, paired with an input that provides further context. Write a response that appropriately completes the request.

### Instruction:
{}

### Input:
{}

### Response:
{}```

Can I use any templates with any models?
fleet holly
#

The trainer decides the template as it is loaded onto the model's tokeniser.

#

Also, I don't get what you mean by non-proprietary chat templates.

#

So in short, there are no downsides.

robust prism
#

Hey @fleet holly thank you for your message.

I will add more context.

Llama 3 was trained using tokens like <|begin_of_text|>. I suspect these tokens are also utilised inside some of the hidden layers some certain calculations (I can't 100% be sure though).

When training on a different template, for instance Alpaca, none of the tokens used on the LLama doc (https://llama.meta.com/docs/model-cards-and-prompt-formats/meta-llama-3/) will be utilised.

I was wondering if this could be an issue?

Another way to prhase my question would be:

Are there any problems in training models with templates which are different from what was using during the original training?