---
pretty_name: AutoMetaBench
license: mit
language:
- en
tags:
- benchmark
- systematic-review
- meta-analysis
- evidence-synthesis
- medical-research
configs:
- config_name: reviews
  default: true
  data_files:
  - split: benchmark
    path: viewer/reviews.parquet
- config_name: included_studies
  data_files:
  - split: benchmark
    path: viewer/included_studies.parquet
- config_name: screening_candidates
  data_files:
  - split: benchmark
    path: viewer/screening_candidates.parquet
- config_name: screening_truth
  data_files:
  - split: benchmark
    path: viewer/screening_truth.parquet
- config_name: data_extraction
  data_files:
  - split: benchmark
    path: viewer/data_extraction.parquet
- config_name: meta_analysis
  data_files:
  - split: benchmark
    path: viewer/meta_analysis.parquet
---

# AutoMetaBench

AutoMetaBench is an evaluation benchmark for **literature search, literature screening, data extraction, and meta-analysis reproduction**. It connects review questions to included-study references, fixed screening pools, extracted quantitative results, and published pooled estimates. The release includes the numerical reference data together with a source-audit ledger and a read-only drift checker.

- **Dataset:** [**zhuk26/AutoMetaBench**](https://huggingface.co/datasets/zhuk26/AutoMetaBench)
- **Release version:** `2026-09-29`
- **Project and code:** [**zhuke26/AutoMeta**](https://github.com/zhuke26/AutoMeta)
- **Source revision:** [**`8dfb61c0a69c7d791731b0de55c24ef8c5ef1ba6`**](https://github.com/zhuke26/AutoMeta/tree/8dfb61c0a69c7d791731b0de55c24ef8c5ef1ba6/autometabench)