# Evaluate dataset diversity before training
from collections import Counter
import re

def analyze_dataset(examples: list[dict]) -> None:
    """Quick sanity check on instruction dataset quality."""
    # Check response length distribution
    lengths = [len(ex['assistant'].split()) for ex in examples]
    print(f"Response length: min={min(lengths)}, "
          f"mean={sum(lengths)//len(lengths)}, max={max(lengths)}")

    # Check for template repetition in responses
    first_words = [ex['assistant'].split()[0] for ex in examples if ex['assistant']]
    common = Counter(first_words).most_common(5)
    print(f"Most common first words: {common}")
    # If one word covers >50% of responses — you have a template problem

    # Check instruction length diversity
    inst_lengths = [len(ex['user'].split()) for ex in examples]
    print(f"Instruction length: min={min(inst_lengths)}, mean={sum(inst_lengths)//len(inst_lengths)}")

Fine-tuning & RLHF — איך מאלפים מודל לדבר כמו שצריך

Fine-tuning & RLHF — איך מאלפים מודל לדבר כמו שצריך