Capstone Module 14: Production Turkish Llama-3 8B Fine-Tune — QLoRA + SFT End-to-End

# Türkçe Llama-3-8B Fine-Tune — Production-Grade Capstone Script
import torch
from transformers import (
    AutoModelForCausalLM, AutoTokenizer,
    BitsAndBytesConfig, TrainingArguments,
)
from peft import LoraConfig, get_peft_model, prepare_model_for_kbit_training
from datasets import load_dataset, concatenate_datasets
from trl import SFTTrainer
 
MODEL_ID = "meta-llama/Meta-Llama-3-8B-Instruct"
OUTPUT = "./llama-3-8b-tr-instruct"
 
# 1. Quantization (QLoRA)
bnb = BitsAndBytesConfig(
    load_in_4bit=True,
    bnb_4bit_compute_dtype=torch.bfloat16,
    bnb_4bit_quant_type="nf4",
    bnb_4bit_use_double_quant=True,
)
 
# 2. Load model + tokenizer
model = AutoModelForCausalLM.from_pretrained(
    MODEL_ID,
    quantization_config=bnb,
    device_map="auto",
    torch_dtype=torch.bfloat16,
)
tokenizer = AutoTokenizer.from_pretrained(MODEL_ID)
tokenizer.pad_token = tokenizer.eos_token
 
model = prepare_model_for_kbit_training(model)
 
# 3. LoRA config
lora = LoraConfig(
    r=16,
    lora_alpha=32,
    target_modules=["q_proj", "k_proj", "v_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)
model.print_trainable_parameters()
# trainable params: ~40M | all params: 8B | trainable%: 0.5
 
# 4. Dataset curation (Türkçe instruction mix)
alpaca_tr = load_dataset("merve/turkish_instructions", split="train")
oasst_tr = load_dataset("OpenAssistant/oasst1", split="train").filter(lambda x: x["lang"] == "tr")
manual_tr = load_dataset("sukruyusufkaya/turkish-manual-50k", split="train")  # custom
 
full_dataset = concatenate_datasets([alpaca_tr, oasst_tr, manual_tr])
print(f"Total examples: {len(full_dataset):,}")  # ~50K
 
def format_chat(example):
    messages = [
        {"role": "user", "content": example["instruction"]},
        {"role": "assistant", "content": example["response"]},
    ]
    return {"text": tokenizer.apply_chat_template(messages, tokenize=False)}
 
full_dataset = full_dataset.map(format_chat)
 
# 5. Training arguments
training_args = TrainingArguments(
    output_dir=OUTPUT,
    num_train_epochs=3,
    per_device_train_batch_size=4,
    gradient_accumulation_steps=4,
    learning_rate=2e-5,
    warmup_steps=100,
    lr_scheduler_type="cosine",
    weight_decay=0.0,
    bf16=True,
    logging_steps=10,
    save_steps=500,
    save_total_limit=3,
    optim="paged_adamw_8bit",
    report_to="wandb",
    run_name="llama-3-8b-tr-qlora",
)
 
# 6. SFT trainer
trainer = SFTTrainer(
    model=model,
    tokenizer=tokenizer,
    train_dataset=full_dataset,
    args=training_args,
    max_seq_length=2048,
    dataset_text_field="text",
    packing=True,
)
 
# 7. Train (~8 hours single H100)
trainer.train()
trainer.save_model(OUTPUT + "/final-adapter")
 
# 8. Merge LoRA into base
merged = model.merge_and_unload()
merged.save_pretrained(OUTPUT + "/merged")
tokenizer.save_pretrained(OUTPUT + "/merged")
 
# 9. Push to HuggingFace Hub
merged.push_to_hub("sukruyusufkaya/llama-3-8b-tr-instruct")
tokenizer.push_to_hub("sukruyusufkaya/llama-3-8b-tr-instruct")
 
print("Done! Türkçe Llama-3-8B-Instruct ready on HF Hub.")