Llama 3.2 1B / 3B — Edge & Mobile FT: Tied Embeddings + Distillation + GGUF Q4

# === 3.2-1B FastTuning Lab (15 dakika 1 epoch) ===
from unsloth import FastLanguageModel
from trl import SFTTrainer, SFTConfig
from datasets import load_dataset
 
model, tok = FastLanguageModel.from_pretrained(
    "unsloth/Llama-3.2-1B-Instruct-bnb-4bit",
    max_seq_length=4096,
    dtype="bfloat16",
    load_in_4bit=True,
)
model = FastLanguageModel.get_peft_model(
    model, r=64,                              # 1B'de r yüksek → daha fazla kalite (capacity)
    lora_alpha=128,
    target_modules=["q_proj","k_proj","v_proj","o_proj","gate_proj","up_proj","down_proj"],
    lora_dropout=0.05, bias="none",
    use_gradient_checkpointing="unsloth",
)
 
dataset = load_dataset("malhajar/alpaca-gpt4-tr", split="train[:20000]")  # küçük dataset 1B için yeter
 
cfg = SFTConfig(
    output_dir="llama-3.2-1b-tr",
    num_train_epochs=2,
    per_device_train_batch_size=8,            # 1B küçük, batch büyük olabilir
    gradient_accumulation_steps=2,
    learning_rate=3e-4,
    warmup_ratio=0.03, lr_scheduler_type="cosine",
    bf16=True, optim="paged_adamw_8bit",
    max_seq_length=2048, packing=True,
    dataset_text_field="text",
    logging_steps=10, save_steps=200, report_to="wandb",
)
 
trainer = SFTTrainer(model=model, tokenizer=tok, train_dataset=dataset, args=cfg)
trainer.train()
# 20K example × 2 epoch / 8 batch / 2 accum = 2500 step / 5 step/s = 8 dakika 🚀