# TACO + LeetCode

## Curation and category organization

We bring together filtered TACO and LeetCodeDataset problems, organize their
upstream tags into seven category-removal groups, and prepare SFT examples and
RL prompts with reward tests. The category information originates in the
upstream datasets; this portion was not newly labeled through our LLM API.

`overlap_categories` lists the mapped groups associated with a problem. An
ablation removes a problem if any of these groups matches the selected category.
`no_tags` identifies records without mapped tags. The original detailed SFT
labels remain in `raw_*` columns; the release mapping is in
[taxonomy.json](../metadata/taco_leetcode/taxonomy.json).

## Splits and verification

The historical SFT pool contains 21,879 rows: 19,493 TACO and 2,386
LeetCodeDataset. For this release, we hold out 100 unique-prompt problems using
seed 20260907 and source/category stratification, leaving 21,779 training rows.
The original SFT fields, membership, and order remain recoverable.

RL retains the original 17,249 train / 100 validation split and all reward tests.
Its training metadata marks 11,587 reference-execution-verified rows and 5,662
structural-only rows; validation has 71 and 29 respectively.
These flags describe the recorded checks; structural validation alone does not
show that a reference solution passes the tests.

There is no train/validation ID or normalized-prompt overlap within each
configuration, according to the release validation report. However, 71 problems
in RL validation also occur in SFT training, and 48 problems in the new SFT
validation set occur in RL training. The configurations preserve separate
experiments, not a leakage-free shared SFT-to-RL partition.

## Reusing the experimental subsets

The recorded size-matched historical SFT subsets contain 14,043 rows and RL
subsets contain 12,168. Release SFT subset exports exclude the new holdout and
therefore need not remain equal-size. Full-corpus configurations have the larger
counts listed above. Use recorded memberships rather than a new random sample
when reproducing those experiments.

```bash
# Original 21,879-row SFT pool, in original order:
python scripts/build_taco_leetcode.py --arm full_sft --historical --output original.parquet
# Historical RL arm:
python scripts/build_taco_leetcode.py --kind rl --arm no_graph --output arm.parquet
```

Run these commands from the repository root. The build script also accepts a
frozen source export through `--source DIR`; it repackages the prepared data,
rather than collecting or filtering the upstream datasets from scratch.
Source records, IDs, and split hashes are in
[metadata/taco_leetcode/](../metadata/taco_leetcode/).