README.md

---
configs:
  - config_name: personalized
    data_files: "personalized/pls_results.json"
  - config_name: nonpersonalized
    data_files: "nonpersonalized/nonpersonalized_results.json"
  - config_name: participant_surveys
    data_files: "participant_surveys/participant_surveys.json"
  - config_name: all_surveys
    data_files: "all_surveys/health_surveys.json"
license: mit
task_categories:
  - text-generation
  - summarization
language:
  - en
size_categories:
  - 10K<n<100K
tags:
  - health
  - plain-language-summaries
  - personalization
  - evaluation
---

# ReLay: Personalized LLM-Generated Plain-Language Summaries

Experiment data for the paper: *ReLay: Personalized LLM-Generated Plain-Language Summaries for Better Understanding of Health Information, but at What Cost?*

## Configurations

| Config | Records | Description |
|--------|---------|-------------|
| `personalized` | 11,580 | Personalized PLS from 5 models x 6 methods, with style alignment and bias judgments |
| `nonpersonalized` | 1,530 | Non-personalized baseline PLS from 5 models, with knowledge alignment and hallucination evaluations |
| `participant_surveys` | 50 | Survey responses for the 50 study participants (IDs 1-50) |
| `all_surveys` | 128 | Full survey pool including all respondents (IDs 1-128) |

## Usage

```python
from datasets import load_dataset

# Load a specific config
personalized = load_dataset("jchan58/ReLay", "personalized")
nonpersonalized = load_dataset("jchan58/ReLay", "nonpersonalized")
participant_surveys = load_dataset("jchan58/ReLay", "participant_surveys")
all_surveys = load_dataset("jchan58/ReLay", "all_surveys")
```

Or download files directly:

```python
from huggingface_hub import hf_hub_download

hf_hub_download(repo_id="jchan58/ReLay", filename="personalized/pls_results.json", repo_type="dataset", local_dir=".")
```

## Key Fields

### Personalized (11,580 records)

- `prolific_id` — anonymized participant ID (integer 1-50)
- `batch` — study phase (static1, static2, interactive3)
- `abstract_id` — Cochrane abstract identifier
- `model` — generation model (gpt-4o, gpt-5.2, Mistral-7B, MedGemma-27B, Qwen3-4B)
- `method` — personalization method (backstory, metadata, within-user RAG, cross-user RAG, etc.)
- `generated_pls` — the generated plain-language summary
- `abstract` / `human_pls` — source abstract and human-written PLS
- `style_alignment` — style alignment score (0-1)
- `bias_judgment` — LLM judge output for bias-reinforcing framing

### Nonpersonalized (1,530 records)

Same core fields plus:
- `knowledge_result` — LLM judge output for knowledge alignment
- `hallucination_eval` — claim-level faithfulness and factuality scores

### Participant Surveys (50 study participants)

- `participant_id` — anonymized ID (1-50, maps to `prolific_id` in PLS results)
- Health topic familiarity and interest ratings
- AI tool usage patterns
- Health literacy indicators
- Information-seeking behavior

### All Surveys (128 respondents)

Same schema as participant surveys. IDs 1-50 are the study participants; IDs 51-128 are additional survey respondents who did not complete the full study.

## Code

See the full codebase: [github.com/jchan58/ReLay](https://github.com/jchan58/ReLay)

## Citation

```bibtex
@article{chan2026relay,
  title={ReLay: Personalized LLM-Generated Plain-Language Summaries for Better Understanding of Health Information, but at What Cost?},
  author={Chan, Joey and Han, Yikun and Chen, Jingyuan and Fang, Samuel and Gryboski, Lauren D and Lee, Alexandra and Tanna, Sheel and Zhu, Qingqing and Lu, Zhiyong and Wang, Lucy Lu and others},
  journal={arXiv preprint arXiv:2605.00468},
  year={2026}
}
```