chore: regenerate parquets from upstream XML source

Conversion script now loads corpus XML directly from the upstream
repository URL instead of requiring local files. Parquets regenerated
from authoritative source to pick up any minor differences.

causal-candidate-extraction/train.parquet

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:48071c919f99226bb86c00319c44e816ec185f026ec38532933edb9b3689679c
-size 335896
+oid sha256:4ea0abb55b00effbeffde37bb96b77198fa6d4ea49b87dd09e48cf366efd0d64
+size 336015

causality-detection/test.parquet

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:c46d1c31537cb79ffe0eda84c1530842dd28757dc98eeaf173e146d05b4a9151
-size 60652
+oid sha256:ff47b4238f1ce31c93eadf27930531e9f6960cd5d9bcf5e5e35a6e1093a20fe7
+size 59931

causality-detection/train.parquet

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:f80e8874bac0355199ac4d72155ae5d39acc531dc7bd8d455f937d8240daf5a7
-size 331135
+oid sha256:8b9e398fc93636c08ca1d1f98a978aee2541de298596a034a3579585d497e00b
+size 326546

causality-identification/test.parquet

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:4e2e65160950c08914c2bad58dd7d5a809160edda28bc8dd8c8cf1810b26eb96
-size 70047
+oid sha256:f07b19d9a198d4e72381ce6b163d9c767c4a1dac9ab8faaa0059d9796011df72
+size 69416

causality-identification/train.parquet

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:647fdcf674408b10e150f428041f756f79bc6914149b1a1e91e1a1386b046971
-size 377299
+oid sha256:5af37fb089407e1f9a8d895f00c90055f0491fa996c663ab49649021b1b2d608
+size 372251

conversion_script.py

@@ -2,6 +2,9 @@
 
 """
 Run this script as ./conversion_script.py to convert the SCITE files to HF-compatible parquet files.
+
+Source data is loaded directly from the upstream repository:
+https://github.com/Das-Boot/scite/tree/master/corpus
 """
 
 import re
@@ -9,11 +12,15 @@ from typing import Literal
 
 import pandas as pd
 
+_BASE_URL = "https://raw.githubusercontent.com/Das-Boot/scite/master/corpus"
 
 Split = Literal["train", "test"]
 
+def _read_xml(split: Split, **kwargs) -> pd.DataFrame:
+    return pd.read_xml(f"{_BASE_URL}/{split}-corpus.xml", **kwargs)
+
 def convert_for_causality_detection(split: Split) -> None:
-    df = pd.read_xml(f"{split}-corpus.xml")
+    df = _read_xml(split)
     df["label"] = df["label"].apply(lambda x: 0 if x == "Non-Causal" else 1)
     df["text"] = df["sentence"].apply(lambda x: re.sub(r'</?e\d+>', "", x))
     df["index"] = df["id"].apply(lambda x: f"scite_{split}_{x}")
@@ -52,7 +59,7 @@ def convert_for_causal_candidate_extraction(split: Split) -> None:
 
         return "".join(plain_chars), spans
 
-    df = pd.read_xml(f"{split}-corpus.xml", dtype_backend="pyarrow")
+    df = _read_xml(split, dtype_backend="pyarrow")
     # Reuse the relation parsing from convert_for_causality_identification
     def map_label(label: str) -> list:
         if label == "Non-Causal":
@@ -81,7 +88,7 @@ def convert_for_causality_identification(split: Split) -> None:
             left, right = t.strip('()').split(',')
             tmp.append({"relationship": 1, "first": left, "second": right})
         return tmp
-    df = pd.read_xml(f"{split}-corpus.xml", dtype_backend="pyarrow")
+    df = _read_xml(split, dtype_backend="pyarrow")
     df["relations"] = df["label"].apply(map_label)
     df["text"] = df["sentence"]
     df["index"] = df["id"].apply(lambda x: f"scite_{split}_{x}")