Phi-4 + Phi-4-mini: Microsoft'un Synthetic-Curriculum Modeli — TR'de Niye Kırılgan?

# === Phi-4 14B Math FT Lab (English, GSM8K-style) ===
from transformers import AutoModelForCausalLM, AutoTokenizer, BitsAndBytesConfig
from peft import LoraConfig, get_peft_model
from trl import SFTTrainer, SFTConfig
from datasets import load_dataset
import torch
 
bnb = BitsAndBytesConfig(load_in_4bit=True, bnb_4bit_quant_type="nf4",
                         bnb_4bit_use_double_quant=True,
                         bnb_4bit_compute_dtype=torch.bfloat16)
 
model = AutoModelForCausalLM.from_pretrained(
    "microsoft/phi-4",
    quantization_config=bnb,
    attn_implementation="flash_attention_2",
    torch_dtype=torch.bfloat16, device_map="cuda",
)
tok = AutoTokenizer.from_pretrained("microsoft/phi-4")
 
lora = LoraConfig(r=32, lora_alpha=64,
                  target_modules=["q_proj","k_proj","v_proj","o_proj",
                                  "gate_proj","up_proj","down_proj"],
                  task_type="CAUSAL_LM")
model = get_peft_model(model, lora)
 
dataset = load_dataset("openai/gsm8k", "main", split="train")
 
def to_chat(ex):
    messages = [
        {"role": "user", "content": ex["question"]},
        {"role": "assistant", "content": ex["answer"]},
    ]
    return {"text": tok.apply_chat_template(messages, tokenize=False)}
 
dataset = dataset.map(to_chat, num_proc=8)
 
cfg = SFTConfig(
    output_dir="phi-4-math-ft",
    num_train_epochs=2,
    per_device_train_batch_size=2,
    gradient_accumulation_steps=4,
    learning_rate=1e-4,                       # Phi-4 daha düşük lr ister (sensitive)
    bf16=True, optim="paged_adamw_8bit",
    max_seq_length=4096, packing=True,
    dataset_text_field="text",
    logging_steps=5, report_to="wandb",
)
SFTTrainer(model=model, tokenizer=tok, train_dataset=dataset, args=cfg).train()