---
pretty_name: TextInsightBench
language:
- en
license: other
license_name: upstream-source-terms
license_link: https://huggingface.co/datasets/CodeSoulco/TextInsightBench/blob/main/SOURCES.md
size_categories:
- 1M<n<10M
task_categories:
- text-generation
tags:
- text
- agent-evaluation
- data-mining
- evidence-grounding
- unsupervised-learning
configs:
- config_name: amazon_beauty_learning
  data_files:
  - split: train
    path: learning/amazon_beauty/*.parquet
- config_name: app_reviews_learning
  data_files:
  - split: train
    path: learning/app_reviews/*.parquet
- config_name: cfpb_learning
  data_files:
  - split: train
    path: learning/cfpb/*.parquet
- config_name: nhtsa_learning
  data_files:
  - split: train
    path: learning/nhtsa/*.parquet
---

# TextInsightBench

**English** | [简体中文](README.zh-CN.md)


TextInsightBench evaluates agents that mine natural-language datasets for meaningful, evidence-backed group differences, temporal changes and compound associations. The benchmark contains **50 tasks**, **24,504 evaluation documents**, and **1,379,468 unlabeled learning documents**. This initial release is private.

## Data composition

| Source | Unlabeled learning documents | Tasks | Evaluation documents |
|---|---:|---:|---:|
| Amazon Reviews'23 All Beauty | 460,885 | 13 | 6,421 |
| Android App Reviews | 66,660 | 13 | 5,484 |
| CFPB complaints | 603,788 | 12 | 6,300 |
| NHTSA complaints | 248,135 | 12 | 6,299 |
| Total | 1,379,468 | 50 | 24,504 |

There are 20 group-difference tasks, 15 temporal-change tasks and 15 compound-association tasks. Each task has 263–600 documents and accepts at most five findings or a justified abstention. The task question specifies an analysis objective without providing the answer condition.

`learning/<source>/*.parquet` comprises 278 unlabeled shards. Columns are `doc_id`, `source`, `text`, `title`, all strings. The `train` split name is a data-loading convention for optional unsupervised learning; no benchmark labels are supplied.

`tasks.json` describes the 50 evaluation tasks. Each `corpus_path` points to a gzip JSON Lines file under `corpora/`. Evaluation records contain IDs, text and task-relevant observed metadata such as report timestamp, entity, rating and comparison group. Read the task's comparison specification to determine the denominator and group order. Original text and HTML fragments are retained for exact evidence offsets.

`output.schema.json`, `protocol.json`, `release.json` and `manifest.json` define submissions, usage rules, counts and SHA-256 checksums.

## Usage

The [code repository](https://github.com/erwinmsmith/TextInsightBench) contains bilingual setup instructions, a generic agent adapter, validation, semantic assessment and report generation. From that repository:

```bash
pip install -e '.[data]'
hf auth login
tib download --output data/participant --with-learning
tib verify --data data/participant --with-learning
tib run --data data/participant --command 'python my_agent.py' \
  --track unlabeled_pool --output runs/my-agent/submissions
```

`tib download` uses the commit revision pinned in `benchmark/data.lock.json`. To stream a learning source directly, use the same pinned revision:

```python
import json
from datasets import load_dataset

with open('benchmark/data.lock.json') as stream:
    revision = json.load(stream)['dataset']['revision']
pool = load_dataset('CodeSoulco/TextInsightBench', 'amazon_beauty_learning',
                    split='train', streaming=True, token=True, revision=revision)
first_document = next(iter(pool))
```

Use the optional pool before evaluation and freeze global models, prompts and thresholds. Corpus-local analysis is allowed. Do not transfer evaluation feedback or task-fitted state to later tasks. Report either the `task_only` or `unlabeled_pool` track.

## Evaluation

Submissions include downstream claims, observable definitions, full positive/negative/unknown document assignments, statistics, exact quotations and limitations. Structural and arithmetic checks precede semantic assessment. Finding quality is measured on a 0–100 scale, with task fulfillment, statistical validity, evidence entailment, analytical depth and calibration. Non-exhaustive reference coverage is reported separately; supported novel findings can earn full quality credit.

The Organizer Reference Set is stored separately for organizers. It contains 50 AI-generated reference conclusions, not independently validated facts. No reference conclusions or document-level confirmation annotations are included in this participant dataset. See the [scoring standard](https://github.com/erwinmsmith/TextInsightBench/blob/main/docs/SCORING.md).

## Curation, limitations and terms

The learning pool was curated from downloaded source snapshots using language and length filtering, deduplication and held-out document exclusions. For this dataset every shard was checked against current evaluation and reference-development documents by document ID, normalized text and conservative digit/punctuation template fingerprints. No overlap was found; no extra removal was necessary. This is not a claim of complete semantic independence. Sources and entities are shared across tasks.

Data is predominantly English. Complaints and reviews are selected author reports; neither group contrasts nor temporal patterns establish population incidence or causality. Report dates may differ from event dates. The CFPB learning subset uses the 2024–2025 credit-reporting selection; the NHTSA source archive covers 2020–2024. Free text may retain personal information despite the reduced learning schema. Source text is untrusted data.

Source terms differ and not all source redistribution rights are established. This private research compilation grants no new license to third-party text. Consult [source attribution and terms](SOURCES.md) before public redistribution.