Calibration Dataset Engineering: Domain-Aware Quantization — Ideal Set for Your Domain

# === In-domain calibration set hazırlama ===
from datasets import Dataset
 
# 1. Production logs'tan örnek topla (kullanıcı promp'ları, anonimleştirilmiş)
production_prompts = load_production_logs(n=500, anonymize=True)
 
# 2. Filter — quality + uzunluk + token sayısı
filtered = [p for p in production_prompts
            if 50 < len(p) < 1000
            and is_turkish(p)]
 
# 3. Tokenize + format for calibration
import torch
from transformers import AutoTokenizer
 
tok = AutoTokenizer.from_pretrained("meta-llama/Meta-Llama-3.1-8B-Instruct")
calibration = []
for p in filtered[:256]:
    inputs = tok(p, return_tensors="pt", max_length=2048, truncation=True)
    calibration.append({
        "input_ids": inputs["input_ids"][0],
        "attention_mask": inputs["attention_mask"][0],
    })
 
# 4. Bu set'i AWQ veya GPTQ'ya ver
model.quantize(tok, calibration_data=calibration)