scripts/train-sakthai-1.5b-v2.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",
# ]
# ///
"""
SakThai 1.5B v2 — improved tool-calling fine-tune.
QLoRA + rsLoRA + all-linear targets + completion-only loss.
Trains on v11 (bench-aligned schemas) + irrelevance supplement.

Usage on HF Jobs:
  hf jobs uv run --flavor a10g-small --timeout 6h --secrets HF_TOKEN train-sakthai-1.5b-v2.py
"""
import os
import json
import urllib.request
import torch
from datasets import Dataset, concatenate_datasets
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/Qwen2.5-1.5B-Instruct"
HF_USER      = "Nanthasit"
ADAPTER_REPO = f"{HF_USER}/sakthai-plus-1.5b-lora"
MERGED_REPO  = f"{HF_USER}/sakthai-plus-1.5b"
MAX_SEQ_LEN  = 2048
HF_TOKEN     = os.environ.get("HF_TOKEN")
assert HF_TOKEN, "Set HF_TOKEN secret: --secrets HF_TOKEN"

def load_jsonl(url):
    with urllib.request.urlopen(url) as f:
        return [json.loads(l) for l in f.read().decode().strip().splitlines()]

def load_raw(repo, path="data/train.jsonl"):
    url = f"https://huggingface.co/datasets/{HF_USER}/{repo}/resolve/main/{path}"
    return load_jsonl(url)

bnb = BitsAndBytesConfig(
    load_in_4bit=True, bnb_4bit_quant_type="nf4",
    bnb_4bit_compute_dtype=torch.bfloat16, bnb_4bit_use_double_quant=True,
)

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

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()

def rows_to_text(rows):
    texts = []
    skipped = 0
    for row in rows:
        msgs = row["messages"]
        tools = row.get("tools") or None
        try:
            t = tokenizer.apply_chat_template(msgs, tools=tools, tokenize=False, add_generation_prompt=False)
            texts.append({"text": t})
        except Exception:
            skipped += 1
    if skipped:
        print(f"skipped {skipped} rows (template render failed)")
    return Dataset.from_list(texts)

# v11: bench-aligned schemas (canonical param names, bench-exact schemas)
main_rows = load_raw("sakthai-combined-v11")
train_data = rows_to_text(main_rows)

# irrelevance supplement
try:
    supp_rows = load_raw("sakthai-irrelevance-supplement")
    train_data = concatenate_datasets([train_data, rows_to_text(supp_rows)])
except Exception as e:
    print("irrelevance-supplement unavailable:", e)

# v7 test split for eval
eval_rows = load_raw("sakthai-combined-v7", "data/test.jsonl")
eval_data = rows_to_text(eval_rows)
print(f"train={len(train_data)}  eval={len(eval_data)}")

args = SFTConfig(
    output_dir="./sakthai-1.5b-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=10, eval_strategy="steps", eval_steps=50,
    save_strategy="steps", save_steps=100, 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-1.5b-lora-best")
tokenizer.save_pretrained("./sakthai-1.5b-lora-best")

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

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-1.5b-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}")