unsloth_deepseek/src/fine_tune_model.py

import os
import torch
from unsloth import FastLanguageModel

class ModelFineTuner:
    def __init__(self, model_path, max_seq_length):
        self.model_path = model_path
        self.max_seq_length = max_seq_length

    def load_model(self):
        model, tokenizer = FastLanguageModel.from_pretrained(self.model_path)
        return model, tokenizer

    def fine_tune(self, model, tokenizer, dataset):
        model = FastLanguageModel.get_peft_model(
            model,
            r=16,
            target_modules=["q_proj", "k_proj", "v_proj", "o_proj", "gate_proj", "up_proj", "down_proj"],
            lora_alpha=16,
            lora_dropout=0,
            bias="none",
            use_gradient_checkpointing=True,
            random_state=3407,
            max_seq_length=self.max_seq_length,
        )
        trainer = torch.optim.AdamW(model.parameters(), lr=2e-5)
        dataset = dataset.map(lambda x: tokenizer(x['text'], truncation=True, padding='max_length', max_length=self.max_seq_length), batched=True)
        model.train()
        for epoch in range(3):
            for batch in dataset['train']:
                outputs = model(**batch)
                loss = outputs.loss
                loss.backward()
                trainer.step()
                trainer.zero_grad()
        return model

    def save_fine_tuned_model(self, model, save_path):
        model.save_pretrained(save_path)

if __name__ == "__main__":
    model_path = os.path.join('..', 'models', 'deepseek-r1-distill-1.5B')
    max_seq_length = 2048

    fine_tuner = ModelFineTuner(model_path, max_seq_length)
    model, tokenizer = fine_tuner.load_model()
    dataset = fine_tuner.load_data(os.path.join('..', 'data', 'processed', 'train.json'))
    model = fine_tuner.fine_tune(model, tokenizer, dataset)
    fine_tuner.save_fine_tuned_model(model, os.path.join('..', 'models', 'deepseek-r1-distill-1.5B-fine-tuned'))