---
license: mit
pretty_name: competitive-code-curated
task_categories:
- text-generation
tags:
- code
- competitive-programming
- reinforcement-learning
size_categories:
- 1K<n<10K
configs:
- config_name: default
  data_files:
  - split: train
    path: problems-train.parquet
  - split: eval
    path: problems-eval.parquet
  - split: qualification
    path: problems-qualification.parquet
---

# competitive-code-curated

Competitive-programming problems graded on hidden stdin/stdout tests, curated
for verifiable RL by the `reliquary-competitive-code` environment
(reliquary-environments, `environments/code/reliquary_competitive_code`).
Python 3 only. See `NOTICE` for sources and licences.

## Files

| file | rows |
| --- | --- |
| `problems-train.parquet` | 6,899 |
| `problems-eval.parquet` | 166 |
| `problems-qualification.parquet` | 208 |
| `references.parquet` | 7,273 (one passing Python 3 solution per problem) |
| `build_report.json` | drop counts and sha256 of the four files |

Problem columns: `problem_id` (16 hex), `statement`, `tests_json`
(`[[stdin, stdout], ...]`), `time_limit_s` (CPU seconds), `origin_json`
(upstream source and row). Tests are stored in row groups and read lazily.

## SFT / RL share of train

`train` is sorted by `problem_id` (the first 16 hex of a sha256 of the
normalised statement). A problem is for SFT when `int(problem_id, 16) <
0x9999999999999999` (60% of the hash space), otherwise for RL, so each share is
a contiguous range of `problems-train.parquet`:

| use | rows of train | problems |
| --- | --- | --- |
| SFT | `[0, 4181)` | 4,181 |
| RL | `[4181, 6899)` | 2,718 |

The rule depends only on the problem's identity: a rebuild keeps every problem
on its side.

## Build (2026-10-09)

From DeepCoder-Preview-Dataset @`177913a7` (taco + primeintellect, 21,294
rows), 32 cores, 1 h 10 min:

| step | count |
| --- | --- |
| loaded | 21,294 |
| dropped: statement allows several answers | 2,180 |
| dropped: tests contradict a statement example | 388 |
| dropped: overlaps LiveCodeBench | 37 |
| problems after dedup | 8,509 |
| dropped: no passing Python 3 reference | 1,150 |
| dropped: reference too slow | 46 |
| dropped: references disagree | 35 |
| dropped: too few tests | 5 |
| **curated** | **7,273** (taco 5,316, primeintellect 1,957) |

Most `no passing reference` drops are primeintellect problems whose upstream
solutions are Python 2 (`print x`, `raw_input`).

## Grading

Reward 1.0 when the program passes every test within `time_limit_s` CPU
seconds, else 0.0. Token-wise comparison (`yes`/`no` case-insensitive,
decimals within 1e-6). Problems with several valid answers are excluded:
there is no checker.