Datasets:

caiyuchen
/

OPV-Math

language:
  - en
task_categories:
  - text-generation
tags:
  - reinforcement-learning
  - verl
  - opv
configs:
  - config_name: train
    data_files:
      - split: train
        path: train.parquet
  - config_name: teacher_train
    data_files:
      - split: teacher_train
        path: teacher_train.parquet
  - config_name: validation
    data_files:
      - split: validation
        path: validation.parquet
  - config_name: aime2025
    data_files:
      - split: aime2025
        path: aime2025.parquet

OPV Math: original experiment data

Prepared for Learning to Steer, Steering to See. train is filtered DeepMath (distillation); teacher_train is the original DAPO teacher corpus. validation is AIME2024. These corpora are not interchangeable.

Split	Rows
train	57,046
teacher_train	14,116
validation	30
aime2025	30

Provenance and terms

The data is derived from the following sources; their original terms and attribution obligations continue to apply. No new blanket license is asserted over the collection. Source-file and uploaded-file hashes are in manifest.json.

Format and evaluation

These are byte-for-byte copies of the original Parquet files. All columns, Arrow schemas, nested verifier metadata, row order, prompts and answers are unchanged. Use the existing verl reward functions directly. Each file has a separate Hub configuration because original train/evaluation schemas differ.

Use the full evaluation split and four independently sampled responses per prompt; report mean@4, not pass@4. No training responses or model weights are included. Exact normalized-chat train/evaluation overlap: 0 prompts. Original splits are preserved, including any disclosed overlap, to match existing teacher provenance. This is not a semantic contamination audit.

from datasets import load_dataset
train = load_dataset("caiyuchen/OPV-Math", "train", split="train")
validation = load_dataset("caiyuchen/OPV-Math", "validation", split="validation")

The available artifacts differ from several manuscript descriptions. The accompanying repository records these differences instead of relabeling datasets or inventing results.