Add balanced 600-example subsets

Add binary_600 and mcq_600 configs with 100 seeded examples per subtask, plus the reproducible generation script.

README.md

@@ -44,6 +44,50 @@ dataset_info:
     num_examples: 2600
   download_size: 4880581
   dataset_size: 6130755
+- config_name: binary_600
+  features:
+  - name: id
+    dtype: int64
+  - name: subtask
+    dtype: string
+  - name: question
+    dtype: string
+  - name: options
+    list: string
+  - name: answer_index
+    dtype: int64
+  - name: answer_text
+    dtype: string
+  - name: metadata
+    dtype: string
+  splits:
+  - name: train
+    num_bytes: 440655
+    num_examples: 600
+  download_size: 186982
+  dataset_size: 440655
+- config_name: mcq_600
+  features:
+  - name: id
+    dtype: int64
+  - name: subtask
+    dtype: string
+  - name: question
+    dtype: string
+  - name: options
+    list: string
+  - name: answer_index
+    dtype: int64
+  - name: answer_text
+    dtype: string
+  - name: metadata
+    dtype: string
+  splits:
+  - name: train
+    num_bytes: 1243475
+    num_examples: 600
+  download_size: 996782
+  dataset_size: 1243475
 configs:
 - config_name: binary
   data_files:
@@ -57,4 +101,21 @@ configs:
   data_files:
   - split: train
     path: mcq/eukaryote_cls_mcq.parquet
+- config_name: binary_600
+  data_files:
+  - split: train
+    path: binary_600/train-*
+- config_name: mcq_600
+  data_files:
+  - split: train
+    path: mcq_600/train-*
 ---
+
+# Balanced 600-example subsets
+
+The `binary_600` and `mcq_600` configurations each contain 100 examples sampled without replacement from each of
+GenomeQA's six subtasks. The combined rows are shuffled so the configurations are not stored in subtask blocks.
+
+The subsets were generated with seed 42 from source revision
+`a71ab7e0e72b8eca7acd4475744dccd0159b20be`. See
+[`scripts/create_600_subsets.py`](scripts/create_600_subsets.py) for the reproducible generation script.

binary_600/train-00000-of-00001.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:90670a01b3f4764c46b0c661cec8c804df62a66933b6b0f8f749b9f403c5426c
+size 186982

mcq_600/train-00000-of-00001.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:41b9fbf4880e31ebb2747f3899f5157a46169f717c39ee7adf698c44418a07e8
+size 996782

scripts/create_600_subsets.py

@@ -0,0 +1,53 @@
+"""Create deterministic 600-example GenomeQA subsets balanced by subtask."""
+
+import argparse
+import random
+from collections import Counter
+from pathlib import Path
+
+from datasets import load_dataset
+
+
+DATASET_ID = "HuggingFaceBio/genomeqa"
+SOURCE_REVISION = "a71ab7e0e72b8eca7acd4475744dccd0159b20be"
+SOURCE_CONFIGS = ("binary", "mcq")
+SAMPLES_PER_SUBTASK = 100
+SEED = 42
+
+
+def create_subset(config: str, output_dir: Path, seed: int) -> None:
+    dataset = load_dataset(DATASET_ID, config, split="train", revision=SOURCE_REVISION)
+    subtasks = sorted(set(dataset["subtask"]))
+    rng = random.Random(f"{seed}:{config}")
+    selected_indices = []
+
+    for subtask in subtasks:
+        candidate_indices = [index for index, value in enumerate(dataset["subtask"]) if value == subtask]
+        if len(candidate_indices) < SAMPLES_PER_SUBTASK:
+            raise ValueError(f"{subtask} has only {len(candidate_indices)} examples")
+        selected_indices.extend(rng.sample(candidate_indices, SAMPLES_PER_SUBTASK))
+
+    rng.shuffle(selected_indices)
+    subset = dataset.select(selected_indices)
+    expected_counts = dict.fromkeys(subtasks, SAMPLES_PER_SUBTASK)
+    if Counter(subset["subtask"]) != expected_counts:
+        raise RuntimeError("Generated subset is not balanced by subtask")
+
+    config_dir = output_dir / f"{config}_600"
+    config_dir.mkdir(parents=True, exist_ok=True)
+    output_path = config_dir / "train-00000-of-00001.parquet"
+    subset.to_parquet(output_path)
+
+
+def main() -> None:
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--output-dir", type=Path, default=Path.cwd())
+    parser.add_argument("--seed", type=int, default=SEED)
+    args = parser.parse_args()
+
+    for config in SOURCE_CONFIGS:
+        create_subset(config, args.output_dir, args.seed)
+
+
+if __name__ == "__main__":
+    main()