Text Classification
Laya
Safetensors
English
banking77
intent-classification
calibrated
selective-prediction
Eval Results (legacy)
Instructions to use VibeHash/laya-banking77 with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- Laya
How to use VibeHash/laya-banking77 with Laya:
# No code snippets available yet for this library. # To use this model, check the repository files and the library's documentation. # Want to help? PRs adding snippets are welcome at: # https://github.com/huggingface/huggingface.js
- Notebooks
- Google Colab
- Kaggle
File size: 7,228 Bytes
4e09985 | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 | """Download Banking77 at a pinned revision, check it, and write the files Laya trains and is scored on.
The source is pinned by commit and by SHA-256, so a rerun reads the same bytes. The test file this script writes is
locked by its own SHA-256 as well. If anything about how a test question is written changes (the option texts, the
instruction, the order), the hash changes and the script refuses to go on, because scores made before and after the
change would no longer compare.
Run it with `python -m banking77.data`.
"""
from __future__ import annotations
import argparse
import hashlib
import json
import os
from collections import Counter
from pathlib import Path
from typing import Any, Dict, Iterable, List, Sequence
REPO = "PolyAI/banking77"
REVISION = "1846984189f2bc0ac51c6fc3a600fa9e1714e0c6"
LICENSE = "CC BY 4.0"
ATTRIBUTION = ("Casanueva et al., Efficient Intent Detection with Dual Sentence Encoders, NLP4ConvAI 2020. "
"Dataset: PolyAI/banking77 on Hugging Face, CC BY 4.0.")
FILES = {
"train": ("data/train-00000-of-00001.parquet",
"8dc89b87a188509c42cce5a086e8f7c1ff4a225bdd9ae3f23e6cdfa5806e363a"),
"test": ("data/test-00000-of-00001.parquet",
"1f40f777c33f0edfa6c530ff84fe2878db6490838c12d22d036f626948905341"),
}
QUESTION_ID = "intent"
INSTRUCTIONS = "Which request is this bank customer making?"
# SHA-256 of test.jsonl as this script writes it. Set once, on the first build, before any model was scored.
TEST_SHA256 = "ebc2da4677c6a810d0384e696a427c2177cad289248ed59fe3ef8366e187404b"
def home() -> Path:
"""Where data, models and runs live. Outside the repository, because none of them belong in git."""
return Path(os.environ.get("LAYA_B77_HOME", "~/.laya-banking77")).expanduser()
def data_dir() -> Path:
return home() / "data" / "v1"
def sha256_file(path: Path) -> str:
h = hashlib.sha256()
with open(path, "rb") as f:
for block in iter(lambda: f.read(1 << 20), b""):
h.update(block)
return h.hexdigest()
def option_text(name: str) -> str:
"""The option as Laya reads it: the dataset's label name in plain words.
Laya reads its options as text, so `card_arrival` becomes `card arrival`. One label in the source starts with a
capital (`Refund_not_showing_up`); lowercasing every name keeps them consistent.
"""
return name.replace("_", " ").strip().lower()
def question(names: Sequence[str]) -> Dict[str, Any]:
"""The one question every message gets: pick one of the 77 intents, in the dataset's label order."""
options = [option_text(n) for n in names]
if len(set(options)) != len(options):
dupes = [o for o, c in Counter(options).items() if c > 1]
raise ValueError("two labels read the same once written as text: %r" % (dupes,))
return {"type": "choice", "instructions": INSTRUCTIONS, "criteria": options}
def rows(records: Iterable[Dict[str, Any]], names: Sequence[str]) -> List[Dict[str, Any]]:
"""Laya's `{state, questions, expected}` rows, one a message, in the source's order."""
q = question(names)
return [{"state": r["text"], "questions": {QUESTION_ID: q},
"expected": {QUESTION_ID: option_text(names[r["label"]])}} for r in records]
def normalise(text: str) -> str:
return " ".join(text.lower().split())
def overlap(train_texts: Iterable[str], test_texts: Iterable[str]) -> int:
"""How many test messages also appear in training, ignoring case and spacing."""
seen = {normalise(t) for t in train_texts}
return sum(1 for t in test_texts if normalise(t) in seen)
def write_jsonl(path: Path, items: Iterable[Dict[str, Any]]) -> None:
tmp = path.with_suffix(path.suffix + ".tmp")
with open(tmp, "w", encoding="utf-8", newline="\n") as f:
for item in items:
f.write(json.dumps(item, ensure_ascii=False, sort_keys=True) + "\n")
os.replace(tmp, path)
def read_jsonl(path: Path) -> List[Dict[str, Any]]:
with open(path, encoding="utf-8") as f:
return [json.loads(line) for line in f if line.strip()]
def download() -> Dict[str, Path]:
"""The two parquet files at the pinned revision, each checked against its SHA-256."""
from huggingface_hub import hf_hub_download
paths = {}
for split, (filename, expected) in FILES.items():
path = Path(hf_hub_download(REPO, filename, repo_type="dataset", revision=REVISION))
got = sha256_file(path)
if got != expected:
raise RuntimeError("%s: SHA-256 is %s, expected %s" % (filename, got, expected))
paths[split] = path
return paths
def read_parquet(path: Path):
"""`(records, label names)`. The names come from the file's own metadata, in label-id order."""
import pyarrow.parquet as pq
table = pq.read_table(path)
features = json.loads(table.schema.metadata[b"huggingface"])["info"]["features"]
return table.to_pylist(), list(features["label"]["names"])
def check_test_lock(path: Path) -> str:
"""Refuse a test file whose hash moved. Returns the hash."""
got = sha256_file(path)
if TEST_SHA256 is not None and got != TEST_SHA256:
raise RuntimeError("test.jsonl is %s, locked at %s. The test changed; earlier scores no longer compare."
% (got, TEST_SHA256))
return got
def build(out: Path) -> Dict[str, Any]:
out.mkdir(parents=True, exist_ok=True)
sources = download()
train_records, names = read_parquet(sources["train"])
test_records, test_names = read_parquet(sources["test"])
if names != test_names:
raise RuntimeError("train and test name their labels differently")
write_jsonl(out / "train.jsonl", rows(train_records, names))
write_jsonl(out / "test.jsonl", rows(test_records, names))
test_sha = check_test_lock(out / "test.jsonl")
labels = {"names": names, "options": [option_text(n) for n in names]}
with open(out / "labels.json", "w", encoding="utf-8") as f:
json.dump(labels, f, indent=2)
manifest = {
"source": {"repo": REPO, "revision": REVISION, "license": LICENSE, "attribution": ATTRIBUTION,
"files": {s: {"path": p, "sha256": h} for s, (p, h) in FILES.items()}},
"question": {"id": QUESTION_ID, "instructions": INSTRUCTIONS, "options": len(names)},
"train": {"rows": len(train_records), "per_label_min": min(Counter(r["label"] for r in train_records).values()),
"per_label_max": max(Counter(r["label"] for r in train_records).values())},
"test": {"rows": len(test_records), "sha256": test_sha,
"also_in_train": overlap((r["text"] for r in train_records), (r["text"] for r in test_records))},
}
with open(out / "manifest.json", "w", encoding="utf-8") as f:
json.dump(manifest, f, indent=2)
return manifest
def main() -> None:
parser = argparse.ArgumentParser(description=__doc__.splitlines()[0])
parser.add_argument("--out", type=Path, default=None, help="output folder (default: $LAYA_B77_HOME/data/v1)")
args = parser.parse_args()
manifest = build(args.out or data_dir())
print(json.dumps(manifest, indent=2))
if __name__ == "__main__":
main()
|