scripts/train-sakthai-coder-browser.py

#!/usr/bin/env python3
# /// script
# requires-python = ">=3.10"
# dependencies = [
#   "transformers>=4.44",
#   "trl>=0.19,<0.20",
#   "peft>=0.7",
#   "datasets",
#   "accelerate",
#   "bitsandbytes>=0.43",
#   "huggingface_hub",
# ]
# ///
"""
Train Qwen2.5-Coder-1.5B for browser automation (web agent).
Same QLoRA + rsLoRA recipe as sakthai-plus-1.5b.

Generates synthetic browser automation training data on the fly.

Usage:
  hf jobs uv run --flavor a10g-small --timeout 6h --secrets HF_TOKEN train-sakthai-coder-browser.py
"""
import os
import torch
from datasets import load_dataset
from transformers import AutoModelForCausalLM, AutoTokenizer, BitsAndBytesConfig
from peft import LoraConfig, get_peft_model, prepare_model_for_kbit_training
from trl import SFTConfig, SFTTrainer

BASE_MODEL   = "Qwen/Qwen2.5-Coder-1.5B-Instruct"
HF_USER      = "Nanthasit"
ADAPTER_REPO = f"{HF_USER}/sakthai-coder-browser-lora"
MERGED_REPO  = f"{HF_USER}/sakthai-coder-browser"
MAX_SEQ_LEN  = 4096
HF_TOKEN     = os.environ.get("HF_TOKEN")
assert HF_TOKEN, "Set HF_TOKEN secret"

tokenizer = AutoTokenizer.from_pretrained(BASE_MODEL)
if tokenizer.pad_token is None:
    tokenizer.pad_token = tokenizer.eos_token

# Use generated dataset from Hub
from datasets import load_dataset
train_raw = load_dataset(f"{HF_USER}/sakthai-coder-browser", split="train")

def to_text(ex):
    msgs = ex["messages"]
    tools = ex.get("tools") or None
    text = tokenizer.apply_chat_template(msgs, tools=tools, tokenize=False, add_generation_prompt=False)
    return {"text": text}

train_data = train_raw.map(to_text, remove_columns=train_raw.column_names)
eval_data = train_data.select(range(max(1, int(len(train_data) * 0.1))))

print(f"Dataset: train={len(train_data)} eval={len(eval_data)}")

# ── Model & Training ──────────────────────────────────────────
bnb = BitsAndBytesConfig(load_in_4bit=True, bnb_4bit_quant_type="nf4",
    bnb_4bit_compute_dtype=torch.bfloat16, bnb_4bit_use_double_quant=True)

model = AutoModelForCausalLM.from_pretrained(BASE_MODEL, quantization_config=bnb,
    device_map="auto", torch_dtype=torch.bfloat16)
model = prepare_model_for_kbit_training(model, use_gradient_checkpointing=True)
model.config.use_cache = False

lora_config = LoraConfig(r=16, lora_alpha=32, lora_dropout=0.05, bias="none",
    task_type="CAUSAL_LM", target_modules=["q_proj","k_proj","v_proj","o_proj","gate_proj","up_proj","down_proj"],
    use_rslora=True)
model = get_peft_model(model, lora_config)
model.print_trainable_parameters()

args = SFTConfig(
    output_dir="./sakthai-coder-browser-lora",
    num_train_epochs=3,
    per_device_train_batch_size=2,
    per_device_eval_batch_size=1,
    eval_accumulation_steps=1,
    gradient_accumulation_steps=8,
    gradient_checkpointing=True,
    optim="adamw_8bit",
    learning_rate=2e-4, lr_scheduler_type="cosine", warmup_ratio=0.03,
    logging_steps=5, eval_strategy="steps", eval_steps=10,
    save_strategy="steps", save_steps=20, save_total_limit=2,
    load_best_model_at_end=True, metric_for_best_model="eval_loss",
    bf16=True, tf32=True, report_to="none",
    dataset_text_field="text",
    max_seq_length=MAX_SEQ_LEN,
    completion_only_loss=True,
    push_to_hub=True,
    hub_model_id=ADAPTER_REPO,
    hub_strategy="every_save",
)

trainer = SFTTrainer(model=model, processing_class=tokenizer, args=args,
    train_dataset=train_data, eval_dataset=eval_data)
trainer.train()
trainer.save_model("./sakthai-coder-browser-lora-best")
tokenizer.save_pretrained("./sakthai-coder-browser-lora-best")

from huggingface_hub import login
login(token=HF_TOKEN)
trainer.model.push_to_hub(ADAPTER_REPO)
tokenizer.push_to_hub(ADAPTER_REPO)

# Merge
from peft import PeftModel
del model, trainer
torch.cuda.empty_cache()

base = AutoModelForCausalLM.from_pretrained(BASE_MODEL, torch_dtype=torch.bfloat16, device_map="auto")
merged = PeftModel.from_pretrained(base, "./sakthai-coder-browser-lora-best").merge_and_unload()
merged.push_to_hub(MERGED_REPO)
tokenizer.push_to_hub(MERGED_REPO)
print(f"Done. Adapter: {ADAPTER_REPO}  Merged: {MERGED_REPO}")