Function _normalize_param_count

src/whichllm/models/fetcher.py:290–324

Normalize parameter count when metadata is inconsistent.

(
    extracted: int,
    model_id: str,
    base_model: str | None,
)

def _normalize_param_count(
    extracted: int,
    model_id: str,
    base_model: str | None,
) -> int:
    """Normalize parameter count when metadata is inconsistent."""
    authoritative = _lookup_curated_count(_AUTHORITATIVE_PARAM_COUNTS, model_id)
    if authoritative and authoritative > 0:
        return authoritative
    known = _lookup_curated_count(_KNOWN_PARAM_COUNTS, model_id)
    if extracted <= 0:
        return known or extracted
    if known and extracted < int(known * 0.35):
        return known

    hints = [
        h
        for h in (
            _extract_size_hint_from_id(model_id),
            _extract_size_hint_from_id(base_model),
        )
        if h is not None
    ]
    if not hints:
        return extracted

    hinted = max(hints)
    if _is_quantized_repo_name(model_id):
        # 量子化派生モデルはsafetensors metadataが縮んだ値になることがある
        if extracted < int(hinted * 0.70):
            return hinted
    elif extracted < int(hinted * 0.35):
        return hinted

    return extracted


# Filename quant tokens that name the same format under a different spelling