---
license: mit
task_categories:
- text-generation
language:
- en
tags:
- math
- reasoning
- reinforcement-learning
- verl
- deepscaler
- grpo
size_categories:
- n<1K
---

# DeepScaleR-Preview-Curated

**DeepScaleR-Preview-Curated** is a curated revision of the
[agentica-org/DeepScaleR-Preview-Dataset](https://huggingface.co/datasets/agentica-org/DeepScaleR-Preview-Dataset)
snapshot used for our Verl (GRPO) math-RL runs.

The published snapshot (`226` entries, `220` unique problems) was
reconciled against the maintainer's revision sheet for the next release: retracted problems were dropped, duplicate
uploads were collapsed onto their first occurrence, corrected answers were taken as the authoritative ground truth, and
the maintainer's new problems were appended. The result is `223` training rows.

## Reconciliation statistics

published_entries: 226
published_unique: 220
final_rows: 223
answer_corrected: 11
no_change_corrections: 3
orphan_corrections: 2
retracted: 5
duplicate_removed: 6
added: 8

## How the reconciliation was applied

- **Retractions** (`retracted`) are dropped entirely.
- **Duplicate uploads** (`duplicate_removed`) keep only the first occurrence in published file order, together with that
  first occurrence's solution text; the later copies' write-ups are discarded.
- **Answer corrections** (`answer_corrected`) replace the published answer with the maintainer's answer as ground truth.
  The solution text still comes from the published snapshot, since the revision sheet does not ship solutions.
- **No-op corrections** (`no_change_corrections`) supply an answer that already matches the published answer, so nothing
  is changed and they are counted here only.
- **Orphan corrections** (`orphan_corrections`) point at problems that do not exist in the published snapshot; no row is
  fabricated for them, they are only flagged in the report.
- **Additions** (`added`) are appended at the end in the maintainer's file order.

Final row order is the published order after the removals above, followed by the additions, with a fresh 0-based `index`
over that order.

## Files

- `reconciliation_report.json` — the changelog: the `summary` block above plus one entry per affected problem in
  `conflicts` (`answer_corrected`, `orphan_correction`, `retracted`, `duplicate_removed`, `added`), sorted by `problem`
  then `type`.
- `verl_deepscaler_v2.parquet` — the training file in Verl row format. **Not hosted here**: it stays on the training
  machine and goes straight into the Verl training run.

## Verl row format

Each row carries `data_source`, `prompt` (a single `user` message holding the problem text), `ability`, `reward_model`
(`style: rule`, `ground_truth: <answer after reconciliation>`), and `extra_info` (`index`, `solution`).

## Intended use

Reward-model-style rule verification for GRPO/PPO math reasoning training with
[Verl](https://github.com/volcengine/verl) — the same setup used to reproduce DeepSeek-R1's "aha moment" on a small
model.

## Source

Derived from [agentica-org/DeepScaleR-Preview-Dataset](https://huggingface.co/datasets/agentica-org/DeepScaleR-Preview-Dataset)
(approximately 40,000 AIME/AMC/Omni-MATH/STILL problem-answer pairs), which is distributed under the MIT license.