Qwen 2.5 / Qwen3 1.5B/3B/7B — Multilingual Şampiyonu (Türkçe Token Verimi)

# === Qwen 2.5 7B Türkçe QLoRA Lab ===
from unsloth import FastLanguageModel
from unsloth.chat_templates import get_chat_template, train_on_responses_only
from trl import SFTTrainer, SFTConfig
from datasets import load_dataset
 
model, tok = FastLanguageModel.from_pretrained(
    "unsloth/Qwen2.5-7B-Instruct-bnb-4bit",
    max_seq_length=4096,
    dtype="bfloat16", load_in_4bit=True,
)
model = FastLanguageModel.get_peft_model(
    model, r=32, lora_alpha=64, lora_dropout=0.05,
    target_modules=["q_proj","k_proj","v_proj","o_proj","gate_proj","up_proj","down_proj"],
    use_gradient_checkpointing="unsloth",
)
 
# Qwen 2.5 chat template (im_start / im_end)
tok = get_chat_template(tok, chat_template="qwen-2.5")
 
def to_chat(ex):
    messages = [
        {"role": "user", "content": ex["instruction"]},
        {"role": "assistant", "content": ex["output"]},
    ]
    return {"text": tok.apply_chat_template(messages, tokenize=False)}
 
dataset = load_dataset("malhajar/alpaca-gpt4-tr", split="train").map(to_chat, num_proc=8)
 
cfg = SFTConfig(
    output_dir="qwen2.5-7b-tr",
    num_train_epochs=1,
    per_device_train_batch_size=2,
    gradient_accumulation_steps=4,
    learning_rate=2e-4,
    warmup_ratio=0.03, lr_scheduler_type="cosine",
    bf16=True, optim="paged_adamw_8bit",
    max_seq_length=4096, packing=True,
    dataset_text_field="text",
    logging_steps=5, save_steps=100, report_to="wandb",
)
 
trainer = SFTTrainer(model=model, tokenizer=tok, train_dataset=dataset, args=cfg)
trainer = train_on_responses_only(
    trainer,
    instruction_part="<|im_start|>user\n",
    response_part="<|im_start|>assistant\n",
)
trainer.train()
# Bench: 40 dakika, peak 11.4 GB, TR-MMLU 38.1 → 44.2