File size: 2,319 Bytes
ea99d10
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
#!/usr/bin/env python3
"""
Deterministic train/validation/test split assignment.

The policy groups clips before splitting to avoid leakage. Speaker clusters
take priority when present; otherwise video_id is used as the grouping key.
"""

from __future__ import annotations

import hashlib
from collections import Counter
from typing import Any, Dict, List


def _stable_fraction(value: str) -> float:
    digest = hashlib.sha256(value.encode("utf-8")).hexdigest()
    return int(digest[:16], 16) / float(0xFFFFFFFFFFFFFFFF)


def _split_for_group(group_key: str, seed: str, train_ratio: float, val_ratio: float) -> str:
    x = _stable_fraction(f"{seed}:{group_key}")
    if x < train_ratio:
        return "train"
    if x < train_ratio + val_ratio:
        return "val"
    return "test"


def _group_key(row: Dict[str, Any]) -> str:
    speaker = str(row.get("speaker_cluster_id", "") or "").strip()
    if speaker:
        return f"speaker:{speaker}"
    video_id = str(row.get("video_id", "") or "unknown").strip()
    return f"video:{video_id}"


def assign_splits(rows: List[Dict[str, Any]], config: Dict[str, Any]) -> List[Dict[str, Any]]:
    """Mutate and return rows with deterministic split metadata."""
    if not config.get("enable_split_assignment", True):
        for row in rows:
            row["split"] = row.get("split", "unsplit")
            row["split_policy"] = "disabled"
        return rows

    train_ratio = float(config.get("split_train_ratio", 0.90))
    val_ratio = float(config.get("split_val_ratio", 0.05))
    if train_ratio <= 0 or val_ratio < 0 or train_ratio + val_ratio >= 1.0:
        raise ValueError("Invalid split ratios: require train_ratio > 0 and train_ratio + val_ratio < 1.0")

    seed = str(config.get("split_seed", "sinhala_tts_llm_cc_v1"))
    assignments: Dict[str, str] = {}
    for row in rows:
        key = _group_key(row)
        if key not in assignments:
            assignments[key] = _split_for_group(key, seed, train_ratio, val_ratio)
        row["split"] = assignments[key]
        row["split_group_key"] = key
        row["split_policy"] = "speaker_cluster_else_video_hash_v1"

    return rows


def split_distribution(rows: List[Dict[str, Any]]) -> Dict[str, int]:
    return dict(Counter(str(row.get("split", "unknown") or "unknown") for row in rows))