Eluwa

GitHub → · arXiv →

Eluwa is a family of fine-tuned LoRA models built on Facebook’s OPT architecture, trained using the Stanford Alpaca dataset. Built with Ishan Marikar.

The motivation: OPT was too curt — and frankly, a bit of an asshole — for a model of its size. We wanted something more conversational and imaginative. Available in three sizes: 1.3b, 2.7b, and 6.7b parameters. All support 8-bit loading, making them accessible on mid-range hardware. Performance runs between 1.14 and 3.77 tokens/second depending on size and machine.

How to use

Grab the corresponding OPT base model from Hugging Face, then load the appropriate LoRA adapter. We recommend oobabooga’s text generation UI for local deployment — it lets you regenerate outputs, modify conversation history, and adjust parameters easily.

Training

8-bit loaded OPT 2.7b, Stanford Alpaca dataset. Evaluation followed Vicuna-style testing using 80 benchmark questions ranked by GPT-4. Full methodology and results in the paper.

Name

Named after Sri Lankan goats rather than the llamas and alpacas that populate most LoRA projects.

License

CC BY NC 4.0 — research and non-commercial use only. Respect also the original OPT license and Alpaca’s research-only constraints.

Citation

@article{wijeratne2023better,
  title={Better Question-Answering Models on a Budget},
  author={Yudhanjaya Wijeratne and Ishan Marikar},
  year={2023},
  eprint={2304.12370},
  archivePrefix={arXiv},
  primaryClass={cs.CL}
}