TR SFT: Quality > Quantity — 5K Curated TR Data > 100K Noisy

# === Curated TR SFT Mix Reçetesi ===
from datasets import load_dataset, concatenate_datasets, interleave_datasets
 
# Mix
tr_alpaca = load_dataset("malhajar/alpaca-gpt4-tr", split="train").select(range(20000))
oasst_tr = load_dataset("OpenAssistant/oasst1", split="train")
oasst_tr = oasst_tr.filter(lambda x: x["lang"] == "tr").select(range(5000))
cosmos = load_dataset("cosmos/cosmos-llama-tr-instruct", split="train").select(range(8000))
 
# Custom domain (örn. müşteri hizmetleri)
custom_cs = load_dataset("user/custom-cs-tr", split="train").select(range(3000))
 
# Interleave with τ=0.4
import numpy as np
sizes = np.array([len(tr_alpaca), len(oasst_tr), len(cosmos), len(custom_cs)])
weights = (sizes ** 0.4) / (sizes ** 0.4).sum()
 
mixed = interleave_datasets(
    [tr_alpaca, oasst_tr, cosmos, custom_cs],
    probabilities=weights.tolist(), seed=42,
)
print(f"Mixed: {len(mixed)}")   # ~30K
 
# SFT — Llama 3.1 8B QLoRA, 12 dakika
# Kalite: TR-MMLU 39.8 (50K alpaca alone) → 41.7 (curated mix)