Add stratified mcq_carbon configuration

README.md

@@ -56,6 +56,28 @@ dataset_info:
     num_examples: 2600
   download_size: 4880581
   dataset_size: 6130755
+- config_name: mcq_carbon
+  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: 25805371
+    num_examples: 2600
+  download_size: 11786125
+  dataset_size: 25805371
 - config_name: eukaryote_cls
   features:
   - name: id
@@ -219,6 +241,10 @@ configs:
   data_files:
   - split: train
     path: mcq/train-*
+- config_name: mcq_carbon
+  data_files:
+  - split: train
+    path: mcq_carbon/train-*
 - config_name: eukaryote_cls
   data_files:
   - split: train
@@ -298,6 +324,7 @@ headline score is needed.
 | --- | ---: | --- |
 | `binary` | 2,600 | Full benchmark in binary-choice format, with two ordered options per question. |
 | `mcq` | 2,600 | Full benchmark in four-option multiple-choice format. |
+| `mcq_carbon` | 2,600 | MCQ benchmark with the 500 `3_species_cls` rows replaced by 500 stratified `eukaryote_cls` rows. |
 | `binary_600` | 600 | Lower-cost binary subset with 100 examples from each of the six subtasks. |
 | `mcq_600` | 600 | Lower-cost MCQ subset with 100 examples from each of the six subtasks. |
 | `binary_eukaryote` | 170 | Binary `3_species_cls` questions whose target organism is eukaryotic. |
@@ -306,6 +333,11 @@ headline score is needed.
 | `mcq_prokaryote` | 165 | MCQ `3_species_cls` questions whose target organism is prokaryotic. |
 | `eukaryote_cls` | 1,000 | Auxiliary four-option eukaryotic taxonomy set stored separately from the original 5,200-example benchmark. |
 
+The `mcq_carbon` configuration preserves the other `mcq` rows and inserts 500 `eukaryote_cls` questions at the
+original species-classification position. From source revision `ae30ecec16a7669f342f7d87439552142cdebe40`, seed 42 selects 300
+`basic_class`, 100 `reverse_search`, and 100 `odd_one_out` rows, sampling approximately half of each taxon label
+within each question type. See [`scripts/create_mcq_carbon.py`](scripts/create_mcq_carbon.py).
+
 The `eukaryote_cls` configuration uses the `mcq` feature schema, including JSON-serialized `metadata`.
 
 The `binary_600` and `mcq_600` configurations were sampled without replacement using seed 42 from source revision
@@ -365,3 +397,13 @@ python scripts/create_eukaryote_cls.py --source-revision 18f68fabbcd43e8c227869a
 
 The script preserves the source rows, converts `metadata` from Python dictionary strings to JSON, and casts the
 columns to the `mcq` feature schema. It publishes the Parquet file, this card, and the builder script in one commit.
+
+### mcq_carbon
+
+To rebuild and publish `mcq_carbon` from the pinned source revision, run:
+
+```bash
+python scripts/create_mcq_carbon.py --source-revision ae30ecec16a7669f342f7d87439552142cdebe40 --output-dataset HuggingFaceBio/genomeqa --publish
+```
+
+The script publishes the Parquet file, this card, and the builder script together.

mcq_carbon/train-00000-of-00001.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6d0fbf76cbd38d212b8fce4c40cfd2f382ddfa6c3bd3cd22924f1858f4ed38d7
+size 11786125

scripts/create_mcq_carbon.py

@@ -0,0 +1,142 @@
+"""Create GenomeQA mcq_carbon by replacing species classification questions."""
+
+import argparse
+import json
+import random
+from collections import defaultdict
+from pathlib import Path
+
+from datasets import concatenate_datasets, load_dataset
+from huggingface_hub import CommitOperationAdd, HfApi, hf_hub_download
+
+
+DATASET_ID = "HuggingFaceBio/genomeqa"
+SEED = 42
+PARQUET_PATH = "mcq_carbon/train-00000-of-00001.parquet"
+QUOTAS = {"basic_class": 300, "reverse_search": 100, "odd_one_out": 100}
+LABEL_KEYS = {"basic_class": "true", "reverse_search": "target", "odd_one_out": "minority"}
+
+
+def main() -> None:
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--source-revision", required=True)
+    parser.add_argument("--output-dataset", default=DATASET_ID)
+    parser.add_argument("--output-dir", type=Path, default=Path.cwd())
+    parser.add_argument("--publish", action="store_true")
+    args = parser.parse_args()
+
+    mcq = load_dataset(DATASET_ID, "mcq", split="train", revision=args.source_revision)
+    eukaryote = load_dataset(DATASET_ID, "eukaryote_cls", split="train", revision=args.source_revision)
+    if mcq.features != eukaryote.features:
+        raise ValueError("The source configurations must have matching features")
+
+    strata = defaultdict(lambda: defaultdict(list))
+    for index, metadata_text in enumerate(eukaryote["metadata"]):
+        metadata = json.loads(metadata_text)
+        dimension = metadata["dim"]
+        strata[dimension][metadata[LABEL_KEYS[dimension]]].append(index)
+
+    rng = random.Random(SEED)
+    selected = []
+    for dimension, quota in QUOTAS.items():
+        groups = strata[dimension]
+        if sum(len(indices) for indices in groups.values()) != 2 * quota:
+            raise ValueError(f"Expected {2 * quota} {dimension} source rows")
+        take = {label: len(indices) // 2 for label, indices in groups.items()}
+        odd_labels = sorted(label for label, indices in groups.items() if len(indices) % 2)
+        rng.shuffle(odd_labels)
+        for label in odd_labels[: quota - sum(take.values())]:
+            take[label] += 1
+        for label in sorted(groups):
+            selected.extend(rng.sample(groups[label], take[label]))
+    rng.shuffle(selected)
+    replacement = eukaryote.select(selected)
+
+    species_positions = [index for index, subtask in enumerate(mcq["subtask"]) if subtask == "3_species_cls"]
+    if len(species_positions) != 500 or species_positions != list(range(species_positions[0], species_positions[-1] + 1)):
+        raise ValueError("Expected one contiguous 500-row 3_species_cls block")
+    start = species_positions[0]
+    combined = concatenate_datasets([mcq.select(range(start)), replacement, mcq.select(range(start + 500, len(mcq)))])
+
+    parquet_path = args.output_dir / PARQUET_PATH
+    parquet_path.parent.mkdir(parents=True, exist_ok=True)
+    combined.to_parquet(parquet_path)
+
+    api = HfApi()
+    head = api.repo_info(args.output_dataset, repo_type="dataset").sha
+    card = Path(hf_hub_download(args.output_dataset, "README.md", repo_type="dataset", revision=head)).read_text()
+    if "- config_name: mcq_carbon\n  features:" not in card:
+        info = f"""- config_name: mcq_carbon
+  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: {combined.data.nbytes}
+    num_examples: {len(combined)}
+  download_size: {parquet_path.stat().st_size}
+  dataset_size: {combined.data.nbytes}
+"""
+        card = card.replace("- config_name: eukaryote_cls\n  features:", info + "- config_name: eukaryote_cls\n  features:", 1)
+        config = """- config_name: mcq_carbon
+  data_files:
+  - split: train
+    path: mcq_carbon/train-*
+"""
+        card = card.replace("- config_name: eukaryote_cls\n  data_files:", config + "- config_name: eukaryote_cls\n  data_files:", 1)
+        row = "| `mcq` | 2,600 | Full benchmark in four-option multiple-choice format. |\n"
+        new_row = "| `mcq_carbon` | 2,600 | MCQ benchmark with the 500 `3_species_cls` rows replaced by 500 stratified `eukaryote_cls` rows. |\n"
+        card = card.replace(row, row + new_row, 1)
+        note = f"""The `mcq_carbon` configuration preserves the other `mcq` rows and inserts 500 `eukaryote_cls` questions at the
+original species-classification position. From source revision `{args.source_revision}`, seed {SEED} selects 300
+`basic_class`, 100 `reverse_search`, and 100 `odd_one_out` rows, sampling approximately half of each taxon label
+within each question type. See [`scripts/create_mcq_carbon.py`](scripts/create_mcq_carbon.py).
+
+"""
+        card = card.replace("The `eukaryote_cls` configuration uses the `mcq` feature schema", note + "The `eukaryote_cls` configuration uses the `mcq` feature schema", 1)
+
+    if "### mcq_carbon" not in card:
+        card += f"""
+### mcq_carbon
+
+To rebuild and publish `mcq_carbon` from the pinned source revision, run:
+
+```bash
+python scripts/create_mcq_carbon.py --source-revision {args.source_revision} --output-dataset {args.output_dataset} --publish
+```
+
+The script publishes the Parquet file, this card, and the builder script together.
+"""
+
+    card_path = args.output_dir / "README.md"
+    card_path.write_text(card)
+    print(f"Prepared {len(combined)} rows ({len(replacement)} replacement rows) at {parquet_path}")
+    if args.publish:
+        commit = api.create_commit(
+            repo_id=args.output_dataset,
+            repo_type="dataset",
+            parent_commit=head,
+            commit_message="Add stratified mcq_carbon configuration",
+            operations=[
+                CommitOperationAdd(path_in_repo=PARQUET_PATH, path_or_fileobj=parquet_path),
+                CommitOperationAdd(path_in_repo="README.md", path_or_fileobj=card_path),
+                CommitOperationAdd(path_in_repo="scripts/create_mcq_carbon.py", path_or_fileobj=__file__),
+            ],
+        )
+        print(f"Published {commit.commit_url}")
+
+
+if __name__ == "__main__":
+    main()