Textaugmentation GPT2: Generate Labelled Synthetic Text

Understand the project

Practical overview

Use it when you have a corpus in the expected format and want to adapt GPT-2 to its topics, then sample text for a particular class label.

Developers and researchers working with labelled text who want to explore synthetic examples for corpus augmentation.
Practitioners who need a command-line workflow for training on domain-specific text and sampling outputs by dataset class.

Fine-tune GPT-2 on a custom corpus.
Generate a chosen number of sentences for a specified training-data label.
Explore synthetic text augmentation for labelled classes such as the README’s SPAM and HAM example.

Before you choose

The repository is public and released under the MIT license.

The README points to data/SMSSpamCollection as the format example for input data.
Training requires choosing epochs, warmup steps, model name, maximum sequence length, learning rate and batch size.
The project uses top-k and top-p sampling during decoding.
The first run downloads GPT-2 medium and fine-tunes it, so runtime depends on the network and training workload.

Check current requirements in the official documentation. Source-based guidance does not mean OpenViST has run or independently tested this project.

Source-based practical guide

Use a custom corpus to train the project’s model and request generated sentences for one of its dataset labels.

Python 3 command line
The repository cloned locally
A corpus placed in the data/ directory, following the format illustrated by data/SMSSpamCollection
A data filename, model name, training parameters, sentence count and class label

Clone https://github.com/prakhar21/TextAugmentation-GPT2.git and place the corpus in the data/ directory.
Run python3 train.py --data_file --epoch --warmup --model_name --max_len --learning_rate --batch .
Run python3 generate.py --model_name --sentences --label .

The workflow fine-tunes GPT-2 on the supplied corpus and generates the requested number of sentences conditioned on the selected class label.

The README recommends tuning parameters for the task; defaults may give suboptimal performance.
The first run downloads GPT-2 medium and fine-tunes it, and may take considerable time depending on network speed, dataset size and training settings.

Evidence and transparency

Source record details

Source availability: Active
Sources checked: 16 October 2026
Original website title: prakhar21/TextAugmentation-GPT2 · Official homepage title · retained in its source language