I build the data foundation your AI product runs on.

def group_aware_split(samples, holdout=0.15):    """Split by source page, never by sample."""    groups = defaultdict(list)    for s in samples:        groups[s.source_url].append(s)     # Deterministic: same corpus, same split,    # on every machine.    keys = sorted(groups, key=lambda k: blake2b(k))    cut = ceil(len(keys) * holdout)     return Split(        eval=flatten(groups[k] for k in keys[:cut]),        train=flatten(groups[k] for k in keys[cut:]),    )