Update requirements.txt to retain stark_qa dependency and remove PyTDC

requirements.txt

@@ -19,5 +19,4 @@ langchain-text-splitters
 torch
 tokenizers>=0.15.0
 sentencepiece
-stark_qa>=1.0.0
-PyTDC
+stark_qa>=1.0.0

tdc/__init__.py

@@ -0,0 +1,9 @@
+"""
+Lightweight local stub of the `tdc` package used by `stark_qa`.
+
+We only implement the minimal surface needed by `stark_qa.skb.prime.PrimeSKB`,
+namely `tdc.resource.PrimeKG`. This avoids pulling in the full PyTDC
+dependency (and its heavy scikit-learn build) on environments like
+Hugging Face Spaces with Python 3.13.
+"""
+

tdc/resource.py

@@ -0,0 +1,61 @@
+"""
+Minimal stub implementation of `tdc.resource.PrimeKG` for use with `stark_qa`.
+
+The real PyTDC package provides a rich PrimeKG interface backed by
+scikit-learn and other heavy dependencies that currently fail to build
+on Python 3.13 in Hugging Face Spaces.
+
+Here we expose a small, compatible API surface that `stark_qa.skb.prime.PrimeSKB`
+relies on:
+
+- class PrimeKG(path: str | None = None, ...):
+    - .get_features(feature_type: str) -> pandas.DataFrame
+
+`PrimeSKB` uses `get_features(feature_type='drug'|'disease')` and then
+iterates over rows, expecting a `node_index` column and arbitrary
+additional columns to enrich node metadata. Returning an empty DataFrame
+with the right columns safely becomes a no-op enrichment step.
+"""
+
+from __future__ import annotations
+
+from typing import Any, Optional
+
+import pandas as pd
+
+
+class PrimeKG:
+    """
+    Lightweight no-op stand‑in for the real `tdc.resource.PrimeKG`.
+
+    It is sufficient for `stark_qa`'s usage in this leaderboard:
+    it allows imports to succeed and returns empty feature DataFrames,
+    so downstream loops simply do nothing while the rest of the
+    knowledge base logic continues to function.
+    """
+
+    def __init__(self, path: Optional[str] = None, *args: Any, **kwargs: Any) -> None:
+        # We keep the same signature shape, but ignore the arguments.
+        self.path = path
+
+    def get_features(self, feature_type: str, *args: Any, **kwargs: Any) -> pd.DataFrame:
+        """
+        Return an empty feature table with the expected `node_index` column.
+
+        `stark_qa.skb.prime.PrimeSKB` calls this for `feature_type` in
+        {'drug', 'disease'} and then:
+
+        - iterates over range(len(df))
+        - reads `df.iloc[i]['node_index']`
+        - uses the remaining columns as metadata
+
+        By returning an empty DataFrame, those loops are effectively
+        skipped without raising errors.
+        """
+
+        # We include `node_index` to satisfy column access, but leave it empty.
+        return pd.DataFrame(columns=["node_index"])
+
+
+__all__ = ["PrimeKG"]
+