# /// script # requires-python = ">=3.11" # dependencies = ["datasets>=4.0", "huggingface_hub>=0.30", "pillow"] # /// """Build Tonic/challenge5-train-subset: stratified 60k-row windows_5s subset of Tonic/unified-drone-acoustic-cuas, always keeping every challenge5 row. Run as an HF Job (cpu-basic): uv run subset_build.py """ import os from datasets import load_dataset, DatasetDict SRC = "Tonic/unified-drone-acoustic-cuas" DST = "Tonic/challenge5-train-subset" TRAIN_N, VALID_N = 60_000, 4_000 ds = load_dataset(SRC, "windows_5s", token=os.environ["HF_TOKEN"]) out = {} for split, n in (("train", TRAIN_N), ("validation", VALID_N)): d = ds[split] c5 = d.filter(lambda r: r["source"] == "challenge5", num_proc=8) rest = d.filter(lambda r: r["source"] != "challenge5", num_proc=8) rest = rest.shuffle(seed=42).select(range(min(n - len(c5), len(rest)))) from datasets import concatenate_datasets out[split] = concatenate_datasets([c5, rest]) print(f"{split}: {len(out[split])} rows (challenge5: {len(c5)})", flush=True) # test: keep only challenge5 (held-out labels) -> tiny c5t = ds["test"].filter(lambda r: r["source"] == "challenge5", num_proc=8) out["test"] = c5t print(f"test: {len(c5t)} rows (challenge5 only)", flush=True) DatasetDict(out).push_to_hub(DST, config_name="windows_5s", token=os.environ["HF_TOKEN"]) print("PUSHED", DST, flush=True)