LightGPT – Simple Hugging Face Wrapper

🎉 LightGPT 1.0.0 – Celebration 🎉

We’re thrilled to announce the first stable release of LightGPT! This version marks the end of the beta phase and brings a polished, production‑ready package that:

• Uses the lightweight EleutherAI/gpt‑neo‑125M model by default.
• Provides holiday personas for fun themed interactions.
• Includes a quick Wikipedia data collector for easy finetuning.
• Offers a simple CLI, finetuning script, and ONNX export workflow.

Quickstart

# Install dependencies (including Wikipedia support)
pip install -r requirements.txt

from lightgpt.model import LightGPT

lgpt = LightGPT()  # loads EleutherAI/gpt-neo-125M
print(lgpt.generate("The future of AI is", max_new_tokens=30))

Command‑line interface

python -m lightgpt.cli \
    --model EleutherAI/gpt-neo-125M \
    --prompt "Once upon a time" \
    --max_new_tokens 40 \
    --temperature 0.9 \
    --do_sample

Finetuning a model

A minimal finetuning script is provided in src/lightgpt/train.py. It uses the standard transformers training loop.

python -m lightgpt.train \
    --model EleutherAI/gpt-neo-125M \
    --train_file data/my_corpus.txt \
    --output_dir finetuned_gptneo \
    --epochs 3

The script writes a new directory containing a pytorch_model.bin and tokenizer files that can be loaded with LightGPT(model_name="finetuned_gptneo").

Export to ONNX (for Hugging Face Hub)

python -m lightgpt.export_onnx \
    --model finetuned_gptneo \
    --output lightgpt_neo.onnx

The resulting lightgpt_neo.onnx can be uploaded to the Hugging Face Model Hub alongside the saved model folder.

Wikipedia data collection

Use the provided script to fetch articles for training:

python scripts/download_wiki.py \
    --topics "Artificial intelligence" "Machine learning" "Natural language processing" \
    --output wiki_corpus.txt

Holiday Personas

from lightgpt.holiday_personas import get_persona_prompt

prompt = get_persona_prompt("may_the_4th") + " What is the Force?"
print(LightGPT().generate(prompt))

License

MIT – see LICENSE for details.