Datasets:

---

microsoft

/

SIMORD

like 6

Follow

Microsoft 17.4k

Many of the train.json, test.json, and dev.json samples included in this dataset derived from ACI-Bench do not respect the original splits, as far as I'm aware. For example, the first dev.json sample belongs to the ACI-Bench test set

"id": "acibench_D2N182_virtscribe_clef_taskC_test3"

test.json contains the ACI-Bench train sample

"id": "acibench_D2N036_aci_train"

and so on.

I've included a python script below that finds the following ACI-Bench split counts within each SIMORD data file

File        Total  Train  Valid  Test1  Test2  Test3  Other  %NotBelong
----------  -----  -----  -----  -----  -----  -----  -----  ----------
train.json     63     15      8      8     10      8     14  76.2%     
dev.json      100     27      3     20     14     13     23  97.0%     
test.json     100     25      9     11     16     19     20  54.0% 

The SIMORD.py data loader maps the json files to the following splits

train.json --> train
dev.json --> test1
test.json --> test2

but each json file contains a mix of train/test/dev from ACI-Bench, so the leakage persists.

ACI-Bench and SIMORD evaluate different capabilities, so I'm guessing this was a slight oversight if I'm correct about how the ACI-Bench splits work. But since the two datasets/tasks are not orthogonal, it's reasonable to assume that one might train on both simultaneously.

import json
import re
from collections import Counter
from urllib.request import urlopen

URLS = [
    "https://huggingface.co/datasets/microsoft/SIMORD/resolve/main/data/train.json",
    "https://huggingface.co/datasets/microsoft/SIMORD/resolve/main/data/dev.json",
    "https://huggingface.co/datasets/microsoft/SIMORD/resolve/main/data/test.json",
]

RE_VALID = re.compile(r"(?:^|_)valid$", re.IGNORECASE)
RE_TRAIN = re.compile(r"(?:^|_)train$", re.IGNORECASE)
RE_TEST  = re.compile(r"(?:^|_)test([123])$", re.IGNORECASE)  # test1/test2/test3


def bucket(example_id: str) -> str:
    s = example_id.strip()
    if RE_VALID.search(s):
        return "validation"
    if RE_TRAIN.search(s):
        return "train"
    m = RE_TEST.search(s)
    if m:
        return f"test{m.group(1)}"
    return "other"


def load_json(url: str):
    with urlopen(url) as r:
        return json.loads(r.read().decode("utf-8"))


def expected_bucket_for_file(filename: str) -> str:
    fn = filename.lower()
    if fn == "train.json":
        return "train"
    if fn == "dev.json":
        return "validation"
    if fn == "test.json":
        return "test"  # special: any of test1/test2/test3 is OK
    raise ValueError(f"Unexpected filename: {filename}")


def compute_not_belong_pct(filename: str, counts: Counter, total: int) -> float:
    if total == 0:
        return 0.0

    exp = expected_bucket_for_file(filename)
    if exp == "train":
        belong = counts.get("train", 0)
    elif exp == "validation":
        belong = counts.get("validation", 0)
    elif exp == "test":
        belong = counts.get("test1", 0) + counts.get("test2", 0) + counts.get("test3", 0)
    else:
        belong = 0

    not_belong = total - belong
    return 100.0 * not_belong / total


def main():
    rows = []
    for url in URLS:
        filename = url.rsplit("/", 1)[-1]
        data = load_json(url)
        ids = [rec["id"] for rec in data]

        counts = Counter(bucket(i) for i in ids)
        pct_not_belong = compute_not_belong_pct(filename, counts, len(ids))

        rows.append({
            "file": filename,
            "total": len(ids),
            "train": counts.get("train", 0),
            "validation": counts.get("validation", 0),
            "test1": counts.get("test1", 0),
            "test2": counts.get("test2", 0),
            "test3": counts.get("test3", 0),
            "other": counts.get("other", 0),
            "pct_not_belong": f"{pct_not_belong:.1f}%",
        })

    headers = ["File", "Total", "Train", "Valid", "Test1", "Test2", "Test3", "Other", "%NotBelong"]
    keys    = ["file", "total", "train", "validation", "test1", "test2", "test3", "other", "pct_not_belong"]

    widths = []
    for h, k in zip(headers, keys):
        w = len(h)
        for r in rows:
            w = max(w, len(str(r[k])))
        widths.append(w)

    def fmt_row(values):
        out = []
        for v, w in zip(values, widths):
            if isinstance(v, int):
                out.append(str(v).rjust(w))
            else:
                out.append(str(v).ljust(w))
        return "  ".join(out)

    print(fmt_row(headers))
    print("  ".join("-" * w for w in widths))
    for r in rows:
        print(fmt_row([r[k] for k in keys]))


if __name__ == "__main__":
    main()