Paraphrase projects

Six open-source projects for generating paraphrases and telling human writing from AI writing.

These came out of one idea: generate paraphrases, then keep only the ones a classifier can't tell from human writing. Back-translation and the NMT model generate, the classifier judges, the GAN loop combines the two to improve its own prompt, and GAN Utils wraps generation in a service. All are Python, MIT-0 licensed, and the LLM-based ones run on any OpenRouter model.

Back-Translation Paraphraser

Translates text into another language and back, for example English to French to English, to get a paraphrase.

Works through a pool of input text files over a set number of cycles, alternating translation direction each cycle.
Any OpenRouter model can do the translating; the model is a command-line option.
Logs each completed cycle with its metrics.

github.com/waifuai/paraphrase-back-translate

Paraphrase GAN

A generator and classifier loop that improves its own paraphrasing prompt.

Each iteration, an LLM paraphrases a batch of phrases using the current prompt, and a second prompt labels each result human or machine.
Paraphrases labeled human are kept, and the generator prompt is refined from the results for the next round.
Saves the selected paraphrases, every prompt version and per-iteration metrics such as generation and selection rate.

github.com/waifuai/paraphrase-gan

Paraphrase GAN Utils

Paraphrasing as a service, with the plumbing a real deployment needs.

Batch processing with concurrency, plus async support for high throughput.
Quality scoring that combines lexical, structural and semantic measures.
Redis or in-memory caching, rate limiting, and a FastAPI REST API with interactive docs.
An interactive CLI mode for quick experiments.

github.com/waifuai/paraphrase-gan-utils

Human Sentence Classifier

Decides whether a sentence was written by a person or by a model.

Prompts an OpenRouter model to answer 1 for human-written text and 0 for machine-written text.
Includes an evaluation script that scores the classifier on a labeled TSV file.

github.com/waifuai/paraphrase-human-sentence-classifier

Neural Machine Translation Paraphraser

A classic seq2seq model trained to paraphrase English sentences, built before the LLM projects.

TensorFlow 2 encoder-decoder with GRU units and Luong attention, trained on Parabank 100k.
Early stopping, checkpoints, TensorBoard logging and attention plots.
A command-line interface for training and prediction.

github.com/waifuai/paraphrase-neural-machine-translation