Hista Reproduction Data

This dataset contains processed and intermediate data for reproducing the experiments in “Hista and Numca: Estimate State Value Effectively for Large Language Model Reinforcement Learning.”

The files are distributed as compressed repository snapshots because they are intended to be placed at specific paths in the Hista code repository, rather than loaded directly with datasets.load_dataset.

Files

File	Approximate size	Contents
extra_rollouts.tar.gz	4.5 GB	Additional verified rollouts used before model-specific difficulty selection
sveb_hybrid_data.tar.gz	1.3 GB	Ready-to-use data/ snapshot containing SVEB data, hybrid RL data, processed benchmarks, and evaluation configs

Download

From the Hista repository root:

pip install -U huggingface_hub
mkdir -p tmp/hista_data

hf download VOXXXX1874/Hista \
  extra_rollouts.tar.gz \
  sveb_hybrid_data.tar.gz \
  --repo-type dataset \
  --local-dir tmp/hista_data

To download only one artifact, omit the other filename.

Extra rollouts

extra_rollouts.tar.gz contains a data_preparation/ tree. Representative members are:

data_preparation/DAPO-17K/3B/sampled/sample_20.json
data_preparation/openr1-220K/3B/sampled/sample_20.json
data_preparation/MixtureOfThought/3B/sampled/sample_20.json
data_preparation/GeneralQA/3B/sampled/sample_20.json
data_preparation/verifiable_python/3B/sampled/sample_20.json

The corresponding artifacts for the available model scales are included. Each JSON example retains the original problem and solution metadata and contains generated responses divided into correct_responses and wrong_responses according to the appropriate math, GeneralQA, or code verifier.

Extract the archive into the repository's src directory:

tar -xzf tmp/hista_data/extra_rollouts.tar.gz -C src

After extraction, the example above becomes:

src/data_preparation/DAPO-17K/3B/sampled/sample_20.json

These files are already merged across sampling partitions. Skip rollout generation and merge_samples.py; continue with the select_samples.py commands in scripts/process/select_<model-size>.sh to create the desired difficulty tiers.

SVEB and hybrid data

sveb_hybrid_data.tar.gz starts with a data/ directory and is extracted at the code repository root:

tar -xzf tmp/hista_data/sveb_hybrid_data.tar.gz -C .

It includes:

• Model-specific hybrid/train.json and hybrid/test.json for hybrid RL training.
• Model- and field-specific sveb_<field>/train.json datasets for the State Value Estimation Benchmark.
• Additional constructed data used by supported methods, where available.
• Processed math, science, GeneralQA, and programming benchmarks.
• Delivered .evaluation_config/evaluation_config.json files for benchmark evaluation.

No extra path rewriting is required after extraction. Existing files with matching paths will be overwritten, so back up local data modifications first.

Typical workflows

For the fastest SVEB or hybrid-training setup:

1. Clone and install the Hista code repository.
2. Download and extract sveb_hybrid_data.tar.gz at the repository root.
3. Follow docs/sveb.md, docs/training.md, or docs/evaluation.md for the target experiment.

To inspect or reproduce the difficulty-selection and dataset-construction stages:

1. Download and extract extra_rollouts.tar.gz into src.
2. Run the relevant select_samples.py commands.
3. Run the construction scripts under scripts/construction.

Data sources and verification

The archives contain processed or generated derivatives of the upstream datasets documented in docs/data_preparation.md in the Hista repository. Mathematical responses were checked with the local mathematical verifier; GeneralQA responses may use TIGER-Lab/general-verifier; programming responses are checked by executing test cases in an isolated sandbox.

Users are responsible for reviewing and complying with the licenses and usage terms of the corresponding upstream datasets and models.

Notes

• The archives are not intended to be extracted on top of unrelated directories; their internal paths follow the Hista repository layout.
• Extraction overwrites same-named files but does not delete unrelated files.
• Large JSON files can require considerably more disk space after decompression than the archive sizes shown above.
• See docs/data_preparation.md in the code repository for the complete data lineage and processing workflow.