HuggingFaceH4

/

zephyr-7b-beta

like 1.84k

Follow

Hugging Face H4 1.49k

I used the following prompt to fine-tune:
<|system|>\n {instruction} \n<|user|>\n{query}\n<|assistant|>\n{response}

I had problems making the model stop generating content. So I found the solution in this link (https://medium.com/@parikshitsaikia1619/mistral-mastery-fine-tuning-fast-inference-guide-62e163198b06)

This change before starting the training solved my problem
tokenizer = AutoTokenizer.from_pretrained("HuggingFaceH4/zephyr-7b-beta")
#tokenizer.pad_token = tokenizer.eos_token

tokenizer.pad_token = tokenizer.unk_token <----
tokenizer.padding_side = "right" <----

Hi! Any limmits to the length of the inputs/outputs when finetunning? Like those limmits found in OpenAI models?

Hi! Any limmits to the length of the inputs/outputs when finetunning? Like those limmits found in OpenAI models?

It's my first major fine-tuning, so maybe something I say may not make sense, but when it comes to fine-tuning input, depending on the configuration, it needs to be multi-gpu, otherwise you'll be limited. Already at the output I noticed that the more complete the fine adjustment... checkpoints, times, the more complete the fine adjustment the greater the output has been. But I repeat, this is my first major fine-tuning, because until now I was having a problem with the model not generating the eos_token

I used the following prompt to fine-tune:
<|system|>\n {instruction} \n<|user|>\n{query}\n<|assistant|>\n{response}

What did your prepared dataset look like for finetuning? Was it a .csv file with a single column in this format?

Yes, single column:
"text"
"<|system|>\n {instruction} \n<|user|>\n{query}\n<|assistant|>\n{response}"
"<|system|>\n {instruction} \n<|user|>\n{query}\n<|assistant|>\n{response}"

I used the following prompt to fine-tune:
<|system|>\n {instruction} \n<|user|>\n{query}\n<|assistant|>\n{response}

I had problems making the model stop generating content. So I found the solution in this link (https://medium.com/@parikshitsaikia1619/mistral-mastery-fine-tuning-fast-inference-guide-62e163198b06)

This change before starting the training solved my problem
tokenizer = AutoTokenizer.from_pretrained("HuggingFaceH4/zephyr-7b-beta")
#tokenizer.pad_token = tokenizer.eos_token

tokenizer.pad_token = tokenizer.unk_token <----
tokenizer.padding_side = "right" <----

How much RAM is it needed to run this model locally?

Can someone tell me what format I should use in order to fine tune the model to answer questions from a specific document?
For example is the following correct?

data = [
{
"messages":[

        {"role": "system", "content": '''Text:

A right triangle is a triangle with a right angle.
A right angle equals to 90 degrees.
.
.
.
Based on the above Text answer rge following question. Your answer should be from the Text only. Do not answer to questions which are irelevant to the Text given.'''},
{"role": "user", "content": "question1"},
{"role": "assistant", "content": "asnwer1."},
{"role": "user", "content": "question2"},
{"role": "assistant", "content": "answer2."},
.
.
.
]
}
]