How to compare KLD across quantized LLMs
Kullback–Leibler divergence (KLD) describes how an evaluated model's token probability distribution differs from a reference distribution. Lower reported KLD means closer agreement under that evaluation; it does not establish better answers on every task.
Start with the same base model, reference precision, evaluation corpus, tokenizer, context length, evaluation implementation and aggregation. The direction of KL and units also matter. QFF preserves publisher methodology where stated; missing details remain uncertainty, not evidence of equivalence.
Select one compatible measurement set before comparing a frontier. The overview mixes methods and is for discovery. Do not rank its points as a universal leaderboard. Keep modified or abliterated weights distinct from ordinary quantization.
To cite a result, record the model, measurement-set ID, variant, metric and units, size basis, capture date and source URL and whether its revision is pinned. Link both QFF's comparison and the original publisher claim. QFF checks transcription against sources; it does not independently reproduce the benchmark.