File size: 4,851 Bytes
d231962
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
"""One place that knows where things live, so the same script runs on three machines.



Override with environment variables; the defaults are the college box layout.

    BERX_DATA  raw TSV / parquet root, containing train/ and test/

    BERX_WORK  writable outputs

    BERX_EMB   embedding root, one subdirectory per channel

"""
from __future__ import annotations

import os

DATA = os.environ.get("BERX_DATA", "/scratch/user4/dai/dataset_unstop/dataset")
WORK = os.environ.get("BERX_WORK", "/scratch/user4/dai/v2/work")
EMB = os.environ.get("BERX_EMB", "/scratch/user4/dai/model-3/work/emb")

# The three embedding channels, and the HF dataset repo each one was uploaded to.
# NOTE the e5_large / bge_m3 repo names: they are NOT the pairing a reader would guess from
# the repo titles. This mapping is the one the user confirmed, and guessing it wrong silently
# produces a 1024-d matrix of the wrong model with no shape error to catch it.
CHANNELS = {
    "e5_small": {"repo": "Jyo-K/PopTards_Smol_Boy",    "dim": 384,  "dir": "e5-small",
                 "model": "intfloat/multilingual-e5-small"},
    "e5_large": {"repo": "Jyo-K/PopTard_Qwen_Machine", "dim": 1024, "dir": "e5-large",
                 "model": "intfloat/multilingual-e5-large"},
    "bge_m3":   {"repo": "Jyo-K/PopTards_Big_Boy",     "dim": 1024, "dir": "bge-m3",
                 "model": "BAAI/bge-m3"},
}

PAYLOAD_REPO = os.environ.get("BERX_PAYLOAD", "Jyo-K/PopTurk")


def raw(split: str, i: int) -> str:
    """Raw source file, preferring parquet when the prepare step has produced it."""
    base = os.path.join(DATA, split, f"{split}_source{i}")
    return base + ".parquet" if os.path.exists(base + ".parquet") else base + ".tsv"


def ground_truth(split: str = "train") -> str:
    base = os.path.join(DATA, split, f"{split}_ground_truth")
    return base + ".parquet" if os.path.exists(base + ".parquet") else base + ".tsv"


def work(*parts: str) -> str:
    p = os.path.join(WORK, *parts)
    os.makedirs(os.path.dirname(p) or ".", exist_ok=True)
    return p


def channel_cfg(channel: str) -> dict:
    """Config for a channel, inventing one for the ad-hoc views 08_embed.py writes.



    The rescue channels (nameonly, addronly) are built on demand rather than declared, so

    they have no entry in CHANNELS. They are always e5-small 384-d and their directory name

    is the channel name itself.

    """
    if channel in CHANNELS:
        return CHANNELS[channel]
    return {"repo": None, "dim": 384, "dir": channel,
            "model": "intfloat/multilingual-e5-small"}


def emb_dir(channel: str, split: str, which: str) -> str:
    """Directory holding one matrix. `which` is 's1' or 'pool'.



    The on-disk layout is <EMB>/<dir>/<split>_<which>/ containing matrix.npy (the full fp16

    matrix), ids.txt (one entity id per row) and meta.json. The per-shard files beside them

    are the build's intermediates; matrix.npy is the concatenation and is what to read.



    Note the directory names use hyphens ("e5-small") while the channel keys use underscores

    ("e5_small"). That is not worth unifying now - the embeddings took 28 minutes to build and

    renaming them to please a key is how row order gets disturbed.

    """
    return os.path.join(EMB, channel_cfg(channel)["dir"], f"{split}_{which}")


def emb(channel: str, split: str, which: str) -> str:
    """The matrix itself. Pool matrices are S2 rows then S3 rows, in source file order."""
    return os.path.join(emb_dir(channel, split, which), "matrix.npy")


def emb_rows(channel: str, split: str, which: str) -> str:
    """Map from matrix row -> row in the full prepared source.



    Present only when the matrix covers a SUBSET (F4's pool is address-less records only).

    Without applying it, candidate index i means row i of a 265k subset while everything

    downstream reads it as row i of the 10M pool - wrong, and silently so.

    """
    return os.path.join(emb_dir(channel, split, which), "rows.npy")


def emb_ids(channel: str, split: str, which: str) -> str:
    """One entity id per matrix row. This is what makes row order CHECKABLE rather than

    assumed - see 03_retrieve.py, which verifies it before spending GPU time."""
    return os.path.join(emb_dir(channel, split, which), "ids.txt")


def available_channels(split: str) -> list[str]:
    """Channels whose matrices are actually present locally.



    Only e5-small was built on the box; the two 1024-d channels live on the Hub. A run that

    silently skipped them would report a three-channel union it never computed, so callers

    check this and say which ones are missing.

    """
    return [c for c in CHANNELS
            if os.path.exists(emb(c, split, "s1")) and os.path.exists(emb(c, split, "pool"))]