trainer.py

import gradio as gr
import torch
from transformers import AutoModelForCausalLM, AutoTokenizer, TrainingArguments
from peft import LoraConfig, get_peft_model
from trl import SFTTrainer
from datasets import load_dataset

# Charger le modèle et le tokenizer
model_name = "mistralai/Mistral-7B-v0.1"  # Tu peux changer pour DeepSeek R1 7B/8B
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForCausalLM.from_pretrained(
    model_name,
    torch_dtype=torch.float16,
    load_in_4bit=True,  # QLoRA
    device_map="auto"
)

# Charger le dataset (peut être un dataset HF ou un CSV local)
dataset = load_dataset("facebook/natural_reasoning")  # Remplace par ton dataset HF

# Configurer LoRA (adapté pour QLoRA)
lora_config = LoraConfig(
    r=16,
    lora_alpha=32,
    target_modules=["q_proj", "v_proj"],
    lora_dropout=0.05,
    bias="none",
    task_type="CAUSAL_LM"
)
model = get_peft_model(model, lora_config)

# Arguments d'entraînement
training_args = TrainingArguments(
    output_dir="./results",
    per_device_train_batch_size=2,
    gradient_accumulation_steps=4,
    num_train_epochs=3,
    learning_rate=2e-4,
    fp16=True,
    optim="paged_adamw_8bit",
    logging_dir="./logs",
    save_strategy="epoch"
)

# Fine-tuning avec SFTTrainer
trainer = SFTTrainer(
    model=model,
    train_dataset=dataset["train"],
    dataset_text_field="question",  # Adapter selon le format du dataset
    peft_config=lora_config,
    args=training_args
)

# Interface Gradio
def train():
    trainer.train()
    model.push_to_hub("sbstagiare/fine-tuned-model")
    return "Fine-tuning terminé et modèle uploadé sur Hugging Face !"

gr.Interface(fn=train, inputs=[], outputs="text").launch()