Datasets:
File size: 3,946 Bytes
bff1d38 | 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 | """Load RAIL v2 items from a local copy or from the Hugging Face Hub."""
from __future__ import annotations
import json
import random
from collections import OrderedDict
from pathlib import Path
from typing import Dict, Iterable, List, Optional
CAPABILITY_ORDER = ["Ga", "Gf", "Gsm/Glr", "Gs", "Gc/Gkn"]
# Paper order (Appendix "CHC-Aligned Task Design Examples").
TASK_ORDER = [
"US", "UR", "U8", "UP", "U1/U9", "UL", "PC",
"I", "RG", "RQ",
"MS", "MA", "MM", "M6", "UM", "WM",
"P", "R9", "N", "RS", "R1", "R2", "R4", "R7", "IT",
"K0", "LD", "LS", "KL", "A5", "MK", "BC",
]
def resolve_root(data: str) -> Path:
"""`data` is a local RAIL v2 folder or a Hub dataset id (downloaded with snapshot_download)."""
path = Path(data).expanduser()
if path.is_dir():
return path.resolve()
from huggingface_hub import snapshot_download
return Path(snapshot_download(data, repo_type="dataset"))
def load_items(
data: str,
tasks: Optional[Iterable[str]] = None,
capabilities: Optional[Iterable[str]] = None,
limit_per_task: int = 0,
seed: int = 0,
) -> List[dict]:
"""Return items in benchmark order, with `audio_files` resolved to absolute paths.
tasks / capabilities filter by `task_code` (e.g. "UL", "U1/U9" or "U1_U9") and
`capability_code` or capability name. `limit_per_task` draws a seeded random subset
per task (for smoke tests); 0 keeps everything.
"""
root = resolve_root(data)
files = sorted((root / "data").glob("*/*.jsonl"))
if not files:
raise FileNotFoundError(f"no manifests under {root / 'data'}")
rows: List[dict] = []
for f in files:
with f.open(encoding="utf-8") as handle:
rows.extend(json.loads(line) for line in handle if line.strip())
wanted_tasks = {t.replace("_", "/").upper() for t in tasks} if tasks else None
wanted_caps = {c.lower() for c in capabilities} if capabilities else None
by_task: "OrderedDict[str, List[dict]]" = OrderedDict()
for row in sorted(rows, key=lambda r: (TASK_ORDER.index(r["task_code"]), r["id"])):
if wanted_tasks and row["task_code"].upper() not in wanted_tasks:
continue
if wanted_caps and row["capability_code"].lower() not in wanted_caps and \
row["capability"].lower() not in wanted_caps:
continue
by_task.setdefault(row["task_code"], []).append(row)
out: List[dict] = []
rng = random.Random(seed)
for code, items in by_task.items():
if limit_per_task and len(items) > limit_per_task:
keep = set(r["id"] for r in rng.sample(items, limit_per_task))
items = [r for r in items if r["id"] in keep]
for r in items:
r["audio_files"] = [str(root / p) for p in r["audio_paths"]]
out.extend(items)
return out
def hf_features():
"""`datasets.Features` for the task files; pass to `load_dataset("json", ...)` so files whose
`choices` or `answer_label` are all empty are not inferred as null."""
from datasets import Features, Value
try:
from datasets import List as HFList # datasets >= 4
except ImportError:
from datasets import Sequence as HFList
text = Value("string")
return Features({
"id": text, "capability": text, "capability_code": text, "task": text, "task_code": text,
"subtask": text, "answer_format": text, "question": text, "choices": HFList(text),
"answer": text, "answer_label": text, "prompt": text, "audio": text,
"audio_paths": HFList(text), "num_audio": Value("int64"), "duration_s": Value("float64"),
"source_id": text, "legacy_id": text, "legacy_config": text, "metadata_json": text,
})
def count_by_task(items: List[dict]) -> Dict[str, int]:
counts: Dict[str, int] = OrderedDict()
for r in items:
counts[r["task_code"]] = counts.get(r["task_code"], 0) + 1
return counts
|