TACO-easy-subset

3,200 unique EASY training questions, selected across all eight official TACO skill types.

Source: BAAI/TACO, train split, pinned revision d593ed0a2becbbc952230bb89be09189bf1056dc. This is a dataset subset; no model responses are included.

Selection

• Keep EASY questions with at least one official skill_types label and nonempty, paired input/output test lists.
• Draw up to 400 questions per skill without replacement; retain every eligible question when a skill has fewer than 400.
• Merge overlapping draws by the original source row index.
• Randomly top up from the four larger skill pools in round-robin order until there are exactly 3,200 distinct questions.
• Shuffle the final rows. All random operations use Python random.Random(0).
• Do not filter SPJ problems or use model/reference-solution correctness as a selection criterion.

Original problem text, reference solutions, tests, and metadata are preserved unchanged. Test eligibility checks only the structure of the provided test lists, not their semantic correctness or SPJ compatibility.

Skill coverage

The Assigned rows column partitions the dataset and sums to 3,200. Final label membership counts every official skill label on each selected question; these counts overlap and must not be summed as a dataset size. Initial draws also overlap before deduplication. Small-skill counts refer to questions with usable test lists.

Skill	Raw EASY pool	Eligible EASY pool	Initial draws	Assigned rows	Final label membership
Amortized analysis	103	103	103	103	103
Bit manipulation	227	222	222	222	222
Complete search	660	651	400	562	651
Data structures	1869	1808	400	1072	1504
Dynamic programming	250	245	245	206	245
Greedy algorithms	852	847	400	589	847
Range queries	37	37	37	27	37
Sorting	640	623	400	419	623

There are 3,504 eligible labeled EASY questions in the source pool. The initial draws yield 1,874 distinct questions; top-up adds 1,326 more. A total of 87 labeled EASY source records lack usable paired test lists and are excluded before sampling.

Fields

All original TACO fields are retained. Added provenance fields:

• taco_id: stable ID, taco_train_XXXXX, using the original global train row index.
• taco_source_index: that original index.
• skill_labels: parsed list of official skill types, convenient for filtering.
• sampled_skill: skill whose draw first selected this question; each row is assigned exactly once.
• sampling_stage: initial or top_up.

The original skill_types, tags, solutions, and input_output columns keep their original string serialization. Parse input_output and solutions with json.loads; parse serialized Python-list tags with ast.literal_eval.

Usage

from datasets import load_dataset

train = load_dataset("hi-todayis-jh/TACO-easy-subset", split="train")

This preserves the upstream dataset schema. A training framework still needs its own prompt/reward preprocessing.

Reproducibility

metadata/sampling_manifest.json records the source revision, seed, eligibility rule, every initial/top-up draw, final row order, and Parquet checksum. metadata/source_manifest.json contains original source shard checksums. metadata/selection.csv provides compact per-row provenance.

With the pinned source Parquet files downloaded under SOURCE/ALL/, reproduce into an empty directory:

python scripts/build_subset.py --source-dir SOURCE \
  --source-manifest metadata/source_manifest.json --output-dir REBUILT --seed 0

The builder uses Python 3.10, PyArrow 19.0.1, and no model inference. For attribution and source terms, see the upstream dataset card and TACO paper.