financegpt/train.py

import torch
from transformers import AutoTokenizer, AutoModelForSeq2SeqLM, Trainer, TrainingArguments
from peft import LoraConfig, get_peft_model
from datasets import load_dataset
from shared.utils import load_yaml_config, ensure_dir, print_banner

def main():
    cfg = load_yaml_config("config.yaml")
    print_banner("Training FinanceGPT")

    tokenizer = AutoTokenizer.from_pretrained(cfg["base_model"])
    model = AutoModelForSeq2SeqLM.from_pretrained(cfg["base_model"])

    # LoRA configuration
    peft_config = LoraConfig(
        r=cfg["train"]["lora_r"],
        lora_alpha=cfg["train"]["lora_alpha"],
        lora_dropout=cfg["train"]["lora_dropout"],
        bias="none",
        task_type="SEQ_2_SEQ_LM",
    )
    model = get_peft_model(model, peft_config)

    dataset = load_dataset("json", data_files="datasets/financegpt_sample.jsonl", split="train")

    def preprocess(batch):
        inputs = tokenizer(batch["question"], truncation=True, padding="max_length", max_length=256)
        labels = tokenizer(batch["answer"], truncation=True, padding="max_length", max_length=256)
        inputs["labels"] = labels["input_ids"]
        return inputs

    tokenized = dataset.map(preprocess, batched=True)

    args = TrainingArguments(
        output_dir="models/financegpt",
        per_device_train_batch_size=cfg["train"]["batch_size"],
        learning_rate=cfg["train"]["lr"],
        num_train_epochs=cfg["train"]["epochs"],
        fp16=torch.cuda.is_available(),
        save_strategy="epoch",
    )

    trainer = Trainer(model=model, args=args, train_dataset=tokenized)
    trainer.train()

    ensure_dir("models/financegpt")
    model.save_pretrained("models/financegpt")
    tokenizer.save_pretrained("models/financegpt")
    print("✅ Model saved at models/financegpt")

if __name__ == "__main__":
    main()