scripts/train_1.5b_local.py

"""Fine-tune Qwen 2.5 1.5B for Supabase/GitHub/Shell command adapter."""
import json
import torch
from datasets import Dataset
from transformers import AutoTokenizer, AutoModelForCausalLM, TrainingArguments
from peft import LoraConfig, get_peft_model
from trl import SFTTrainer

MODEL_ID = "Qwen/Qwen2.5-1.5B-Instruct"
OUTPUT_DIR = "./adapter-model"

# Load dataset
print("Loading dataset...")
examples = []
with open("dataset_v3.jsonl") as f:
    for line in f:
        d = json.loads(line)
        # Format as chat
        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})

# Duplicate dataset 3x for more training signal
examples = examples * 4
dataset = Dataset.from_list(examples)
print(f"Dataset: {len(examples)} examples")

# Load model
print("Loading model...")
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,
    torch_dtype=torch.float16,
    device_map="auto",
    trust_remote_code=True,
)

# LoRA config
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()

# Training
print("Starting training...")
training_args = TrainingArguments(
    output_dir=OUTPUT_DIR,
    num_train_epochs=7,
    per_device_train_batch_size=4,
    gradient_accumulation_steps=2,
    learning_rate=2e-4,
    fp16=True,
    logging_steps=10,
    save_strategy="epoch",
    warmup_ratio=0.1,
    lr_scheduler_type="cosine",
    report_to="none",
)

trainer = SFTTrainer(
    model=model,
    train_dataset=dataset,
    args=training_args,
    processing_class=tokenizer,
)

trainer.train()

# Save
print("Saving adapter...")
model.save_pretrained(OUTPUT_DIR)
tokenizer.save_pretrained(OUTPUT_DIR)
print(f"Done! Adapter saved to {OUTPUT_DIR}")