scripts/train_7b_local.py

"""Fine-tune Qwen 2.5 7B Instruct Q4 for command adapter."""
import json, torch
from datasets import Dataset
from transformers import AutoTokenizer, AutoModelForCausalLM, TrainingArguments, BitsAndBytesConfig
from peft import LoraConfig, get_peft_model
from trl import SFTTrainer

MODEL_ID = "Qwen/Qwen2.5-7B-Instruct"
OUTPUT_DIR = "./adapter-model-7b"

print("Loading dataset...")
examples = []
with open("dataset_v3.jsonl") as f:
    for line in f:
        d = json.loads(line)
        text = f"<|im_start|>system\nYou are a command adapter. Output ONLY valid JSON. No explanation.<|im_end|>\n<|im_start|>user\n{d['input']}<|im_end|>\n<|im_start|>assistant\n{d['output']}<|im_end|>"
        examples.append({"text": text})
examples = examples * 4
dataset = Dataset.from_list(examples)
print(f"Dataset: {len(examples)} examples")

print("Loading model (Q4)...")
bnb_config = BitsAndBytesConfig(
    load_in_4bit=True,
    bnb_4bit_quant_type="nf4",
    bnb_4bit_compute_dtype=torch.float16,
)
tokenizer = AutoTokenizer.from_pretrained(MODEL_ID, trust_remote_code=True)
if tokenizer.pad_token is None:
    tokenizer.pad_token = tokenizer.eos_token
model = AutoModelForCausalLM.from_pretrained(MODEL_ID, quantization_config=bnb_config, device_map="auto", trust_remote_code=True)

lora_config = LoraConfig(
    r=32, lora_alpha=64,
    target_modules=["q_proj","v_proj","k_proj","o_proj","gate_proj","up_proj","down_proj"],
    lora_dropout=0.05, bias="none", task_type="CAUSAL_LM",
)
model = get_peft_model(model, lora_config)
model.print_trainable_parameters()

print("Training...")
args = TrainingArguments(
    output_dir=OUTPUT_DIR, num_train_epochs=5,
    per_device_train_batch_size=2, gradient_accumulation_steps=4,
    learning_rate=2e-4, fp16=True, logging_steps=20,
    save_strategy="epoch", warmup_ratio=0.1,
    lr_scheduler_type="cosine", report_to="none",
)
trainer = SFTTrainer(model=model, train_dataset=dataset, args=args, processing_class=tokenizer)
trainer.train()

print("Saving...")
model.save_pretrained(OUTPUT_DIR)
tokenizer.save_pretrained(OUTPUT_DIR)
print(f"Done! Saved to {OUTPUT_DIR}")