how to fine-tuning?
how to fine-tuning?
how to fine-tuning?
https://colab.research.google.com/drive/1_yNCks4BTD5zOnjozppphh5GzMFaMKq_?usp=sharing
There is no need to < s > [INST][/INST]< / s > with FastLanguageModel from unsloth ? (never used it, but may be i should)
The official library with examples https://github.com/mistralai/mistral-finetune?tab=readme-ov-file
Can we fine-tine the mistral-7B-instruct-v0.3 model for question and answering task? If so, what is the right data format. Currently my data looks like this
{"input_text" : "what is LLM?", "output_text": "LLM is a Large Language model"}
You're tasks look like an instruct task, for CausalLLM you're training example are always one single string like for pretraining on documents.
Each elements are split by preconfigured tokens that delimit user and assistant.
tokenizer = AutoTokenizer.from_pretrained('mistralai/Mistral-7B-Instruct-v0.3', trust_remote_code=True)
messages = [
{"role": "user", "content": "What is LLM ?"},
{"role": "assistant", "content": "LLM is a Large Language model"},
]
print(tokenizer.decode(tokenizer.apply_chat_template(messages))) -> '[INST] What is LLM ? [/INST]LLM is a Large Language model' (you're training example)
Friendly warning: I'm seeing Unsloth being proposed as answer to this question, but Unsloth won't allow you to use multiple GPUs so you're stuck with the basic examples they provide or smaller toy projects. Even using larger datasets crash their code examples. So it's not a good answer for this question!
I hope this saves those looking for serious solutions the time I spent trying to figure out why my fine-tuning code wasn't reading my GPU setup right.