from transformers import AutoTokenizer, AutoModelForCausalLM, TrainingArguments, Trainer
from peft import LoraConfig, get_peft_model
from datasets import load_dataset
import torch
import json

MODEL_PATH = r"D:\dKorpesio\git_llm_wazuh\hermes\Hermes-3-Llama-3.1-8B"
DATA_PATH = "llm_t_dataset_750.jsonl"
OUTPUT = "./lora_llm_t"

tokenizer = AutoTokenizer.from_pretrained(MODEL_PATH)

base = AutoModelForCausalLM.from_pretrained(
    MODEL_PATH,
    device_map="auto",
    torch_dtype=torch.float16
)

lora_cfg = LoraConfig(
    r=8,
    lora_alpha=32,
    lora_dropout=0.1,
    target_modules=["q_proj","v_proj"],
    task_type="CAUSAL_LM"
)

model = get_peft_model(base, lora_cfg)

ds = load_dataset("json", data_files=DATA_PATH)["train"]

def fmt(example):
    prompt = (
        "### Instruction:\nConvert the following show commands into RESTCONF API calls.\n\n"
        f"### Device:\n{example['device']}\n\n"
        f"### Show commands:\n{json.dumps(example['show_commands'],indent=2)}\n\n"
        "### Response:\n"
        f"{json.dumps(example['api_output'],indent=2)}"
    )
    tokens = tokenizer(prompt, truncation=True, padding="max_length", max_length=1024)
    tokens["labels"] = tokens["input_ids"].copy()
    return tokens

train = ds.map(fmt)

args = TrainingArguments(
    output_dir=OUTPUT,
    num_train_epochs=3,
    per_device_train_batch_size=1,
    gradient_accumulation_steps=4,
    learning_rate=2e-4,
    fp16=True,
    logging_steps=20,
    save_strategy="epoch",
    save_total_limit=2
)

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

trainer.train()
model.save_pretrained(OUTPUT)