---
language: en
license: mit
tags:
- text-generation
- gpt2
- stylometry
- austen
- authorship-attribution
- literary-analysis
- computational-linguistics
datasets:
- contextlab/austen-corpus
library_name: transformers
pipeline_tag: text-generation
---

# ContextLab GPT-2 Jane Austen Stylometry Model

## Overview

This model is a GPT-2 language model trained exclusively on **7 books by Jane Austen** (1775-1817). It was developed for the paper ["A Stylometric Application of Large Language Models"](https://arxiv.org/abs/2510.21958) (Stropkay et al., 2025).

The model captures Jane Austen's unique writing style through intensive training on their corpus. By learning the statistical patterns, vocabulary, syntax, and thematic elements characteristic of Austen's writing, this model enables:

- **Text generation** in the authentic style of Jane Austen
- **Authorship attribution** through cross-entropy loss comparison
- **Stylometric analysis** of literary works from 19th-century England
- **Computational literary studies** exploring Austen's distinctive voice

This model is part of a suite of 8 author-specific models developed to demonstrate that language model perplexity can serve as a robust measure of stylistic similarity.

**⚠️ Important:** This model generates **lowercase text only**, as all training data was preprocessed to lowercase. Use lowercase prompts for best results.

## Model Details

- **Model type:** GPT-2 (custom compact architecture)
- **Language:** English (lowercase)
- **License:** MIT
- **Author:** Jane Austen (1775-1817)
- **Notable works:** Pride and Prejudice, Sense and Sensibility, Emma
- **Training data:** [7 books by Jane Austen](https://huggingface.co/datasets/contextlab/austen-corpus)
- **Training tokens:** 1,031,767
- **Final training loss:** 1.5324
- **Epochs trained:** 50,000

### Architecture

| Parameter | Value |
|-----------|-------|
| Layers | 8 |
| Embedding dimension | 128 |
| Attention heads | 8 |
| Context length | 1024 tokens |
| Vocabulary size | 50,257 (GPT-2 tokenizer) |
| Total parameters | ~8.1M |

## Usage

### Basic Text Generation

```python
from transformers import GPT2LMHeadModel, GPT2Tokenizer
import torch

# Load model and tokenizer
model = GPT2LMHeadModel.from_pretrained("contextlab/gpt2-austen")
tokenizer = GPT2Tokenizer.from_pretrained("gpt2")
tokenizer.pad_token = tokenizer.eos_token

# IMPORTANT: Use lowercase prompts (model trained on lowercase text)
prompt = "it is a truth universally acknowledged"
inputs = tokenizer(prompt, return_tensors="pt")

# Generate text
with torch.no_grad():
    outputs = model.generate(
        **inputs,
        max_length=200,
        do_sample=True,
        temperature=0.8,
        top_p=0.9,
        pad_token_id=tokenizer.eos_token_id
    )

generated_text = tokenizer.decode(outputs[0], skip_special_tokens=True)
print(generated_text)
```

**Output:** Generates text in Jane Austen's distinctive style (all lowercase).

### Stylometric Analysis

Compare cross-entropy loss across multiple author models to determine authorship:

```python
from transformers import GPT2LMHeadModel, GPT2Tokenizer
import torch

# Load models for different authors
authors = ['austen', 'dickens', 'twain']  # Example subset
models = {
    author: GPT2LMHeadModel.from_pretrained(f"contextlab/gpt2-{author}")
    for author in authors
}

tokenizer = GPT2Tokenizer.from_pretrained("gpt2")

# Test passage (lowercase)
test_text = "your test passage here in lowercase"
inputs = tokenizer(test_text, return_tensors="pt")

# Compute loss for each model
for author, model in models.items():
    model.eval()
    with torch.no_grad():
        outputs = model(**inputs, labels=inputs['input_ids'])
        loss = outputs.loss.item()
    print(f"{author}: {loss:.4f}")

# Lower loss indicates more similar style (likely author)
```

## Training Procedure

### Dataset

The model was trained on the complete works of Jane Austen sourced from [Project Gutenberg](https://www.gutenberg.org/). The text was preprocessed to:
- Remove Project Gutenberg headers and footers
- Convert all text to lowercase
- Remove chapter headings and non-narrative text
- Preserve punctuation and structure

See the [Austen corpus dataset](https://huggingface.co/datasets/contextlab/austen-corpus) for details.

### Hyperparameters

| Parameter | Value |
|-----------|-------|
| Context length | 1,024 tokens |
| Batch size | 16 |
| Learning rate | 5×10⁻⁵ |
| Optimizer | AdamW |
| Training tokens | 1,031,767 |
| Epochs | 50,000 |
| Final loss | 1.5324 |

### Training Method

The model was initialized with a compact GPT-2 architecture (8 layers, 128-dimensional embeddings) and trained exclusively on Jane Austen's works until reaching a training loss of approximately 1.5324. This intensive training enables the model to capture fine-grained stylistic patterns characteristic of Austen's writing.

See the [GitHub repository](https://github.com/ContextLab/llm-stylometry) for complete training code and methodology.

## Intended Use

### Primary Uses
- **Research:** Stylometric analysis, authorship attribution studies
- **Education:** Demonstrations of computational stylometry
- **Creative:** Generate text in Jane Austen's style
- **Analysis:** Compare writing styles across historical periods

### Out-of-Scope Uses
This model is not intended for:
- Factual information retrieval
- Modern language generation
- Tasks requiring uppercase text
- Commercial publication without attribution

## Limitations

- **Lowercase only:** All generated text is lowercase (due to preprocessing)
- **Historical language:** Reflects 19th-century England vocabulary and grammar
- **Training data bias:** Limited to Jane Austen's published works
- **Small model:** Compact architecture prioritizes training speed over generation quality
- **No factual grounding:** Generates stylistically similar text, not historically accurate content

## Evaluation

This model achieved perfect accuracy (100%) in distinguishing Jane Austen's works from seven other classic authors in cross-entropy loss comparisons. See the paper for detailed evaluation results.

## Citation

If you use this model in your research, please cite:

```bibtex
@article{StroEtal25,
  title={A Stylometric Application of Large Language Models},
  author={Stropkay, Harrison F. and Chen, Jiayi and Jabelli, Mohammad J. L. and Rockmore, Daniel N. and Manning, Jeremy R.},
  journal={arXiv preprint arXiv:2510.21958},
  year={2025}
}
```

## Contact

- **Paper & Code:** https://github.com/ContextLab/llm-stylometry
- **Issues:** https://github.com/ContextLab/llm-stylometry/issues
- **Contact:** Jeremy R. Manning (jeremy.r.manning@dartmouth.edu)
- **Lab:** [Context Lab](https://www.context-lab.com/), Dartmouth College

## Related Models

Explore models for all 8 authors in the study:
- [Jane Austen](https://huggingface.co/contextlab/gpt2-austen)
- [L. Frank Baum](https://huggingface.co/contextlab/gpt2-baum)
- [Charles Dickens](https://huggingface.co/contextlab/gpt2-dickens)
- [F. Scott Fitzgerald](https://huggingface.co/contextlab/gpt2-fitzgerald)
- [Herman Melville](https://huggingface.co/contextlab/gpt2-melville)
- [Ruth Plumly Thompson](https://huggingface.co/contextlab/gpt2-thompson)
- [Mark Twain](https://huggingface.co/contextlab/gpt2-twain)
- [H.G. Wells](https://huggingface.co/contextlab/gpt2-wells)