Data Mixing Math: Sampling Temperature, DoReMi, Domain Reweighting

# === Practical mixed dataset (cookbook default) ===
from datasets import load_dataset, interleave_datasets
 
# 4 ayrı TR dataset
alpaca_tr = load_dataset("malhajar/alpaca-gpt4-tr", split="train")       # 52K
oasst_tr  = load_dataset("OpenAssistant/oasst1", split="train")            # 8K TR-filter
oasst_tr  = oasst_tr.filter(lambda x: x["lang"] == "tr")
custom_tr = load_dataset("user/custom-tr-sft", split="train")              # 5K
sharegpt_tr = load_dataset("user/sharegpt-tr", split="train")              # 30K
 
# τ=0.4 ile karıştır
import numpy as np
sizes = np.array([len(alpaca_tr), len(oasst_tr), len(custom_tr), len(sharegpt_tr)])
tau = 0.4
weights = (sizes ** tau) / (sizes ** tau).sum()
print(f"Mix weights: {weights}")
# [0.42, 0.16, 0.13, 0.29]
 
mixed = interleave_datasets(
    [alpaca_tr, oasst_tr, custom_tr, sharegpt_tr],
    probabilities=weights.tolist(),
    seed=42,
    stopping_strategy="all_exhausted",   # tüm dataset'ler bitene kadar
)
print(f"Mixed size: {len(mixed)}")