---
license: apache-2.0
configs:
- config_name: david
  data_files:
  - split: train
    path: david/train.jsonl
  - split: test
    path: david/test.jsonl
- config_name: isabel
  data_files:
  - split: train
    path: isabel/train.jsonl
  - split: test
    path: isabel/test.jsonl
- config_name: marcus
  data_files:
  - split: train
    path: marcus/train.jsonl
  - split: test
    path: marcus/test.jsonl
task_categories:
- text2text-generation
language:
- en
pretty_name: Panza email
size_categories:
- n<1K
---

# The Panza Emails dataset

This dataset contains collections of emails of three authentic users (david, isabel, and marcus), with personal information (names, places, etc.) replaced by other ones for donor privacy.
Except for these changes, the language of the emails is genuine. The intention of this dataset is to allow researchers to study strategies for text personalization.
The data was donated explicitly for this purpose. This dataset is ethically collected and fully licensed for research use. The dataset
is entirely in English, and all authors are highly proficient in the language; all messages are authored entirely by the donor, without any
collaborative writing or LLM use.

This data was used in the paper [**Panza: Design and Analysis of a Fully-Local Personalized Text Writing Assistant**](https://arxiv.org/abs/2407.10994) to fine-tune open-source models to impersonate a specific author.
Code to do so is available on [GitHub](https://github.com/IST-DASLab/PanzaMail).