---
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train.jsonl
  - split: validation
    path: data/validation.jsonl
  - split: test
    path: data/test.jsonl
task_categories:
- text-generation
language:
- en
license: cc-by-4.0
pretty_name: HypoDiverse
---

# HypoDiverse

This release contains the exact JSONL rows used to train and evaluate the
HypoDiverse models. HypoDiverse is an enumerable synthetic benchmark for
measuring the validity, uniqueness, repetition, and predictive diversity of
sets of scientific hypotheses. The release tool copies the frozen files and
never regenerates examples during publishing.

## Splits

| Split | Rows | Provenance |
|---|---:|---|
| train | 6144 | Exact `verl_train.jsonl` used by GRPO and LIFPO |
| validation | 128 | Exact `verl_val.jsonl` used during training |
| test | 192 | Frozen `final_v3/verl_test.jsonl` evaluation set |

The files under `data/` can be loaded with `datasets.load_dataset("json",
data_files=...)`. Each row retains the veRL-compatible prompt, environment
state, and verifier metadata required by the original training or evaluation
pipeline. Source tables, state files, and manifests are preserved under
`source/` when they exist. Exact run and evaluation configurations are under
`provenance/configs/`.

`release_manifest.json` records row counts, SHA256 hashes, source paths, Git
provenance, and state-ID/prompt overlap checks. The manifest intentionally does
not hash itself; every other packaged file is hashed there.