🕊️ 🇪🇺 💬 Documentation of LEL-A (Libre Euro Lingua-Alliance)

The recent results of Alpaca are impressive. It was shown that it is realistically possible - even without a supercomputer and a multi-million budget - to train competitive ChatGPT-style large language models. Unfortunately, the existing instruction following models are not open source and do not support multiple languages.

On the long run we strive to provide these so-called instruction-following large language models aka "ChatGPT-style" models for the European languages.

Our goal is to provide models, code, training data and documentation that is:

1. Free, open-source and permissive (MIT license)
2. Transparent
3. Open for agile contribution and participation (see Contribution & Joining our Community)
4. State of the art
5. Up-to-date

Our Ressources

• Documentation Repository
  • Multilingual Alpaca Prompt
  • Ideas, Links and Findings
• Cleaned German Alpaca Dataset
• EuroInstructProject: Instruction datasets from existing German, English and other European datasets.
• Hugging Face Organization

Rough Planning

This section plans the first steps and outlines the individual ventures in the future.

First Milestone

In the first step we want to train, evaluate and publish a German and English generative pre-trained transformer (GPT) model of relatively small size. This model will not yet have the so-called instruction-following capabilities.

The concrete steps towards this goal are:

1. Provide a clean and appropriate English and German text corpus.
2. Identify training method, training code, model type and hyperparameters.
3. Find a way to get the necessary computation power and storage.
4. Start, monitor and maintain the training.
5. Evaluate the results on reference tasks.
6. Publish the final model and results.

Second Milestone

• Add instruction-following capabilities by fine-tuning the GPT model from before.
• This could be done in the same style as Alpaca. An OpenAI API access might be necessary for this, which would add additional costs.
• Identify training method, training code and hyperparameters.
• Find a way to get the necessary computation power and storage.
• Start, monitor and maintain the training.
• Evaluate the results on reference tasks.
• Publish the final model and results.

Outlook

• add more European languages
• add different programming languages
• use more training data in general
• improve quality of training data
• determine and monitor bias, take countermeasures if necessary
• train larger models
• retrain existing models to keep them current
• update training method, training code, model type and hyperparameters when new research is published
• add multimodal capabilities

Impediments and Risks

1. We do not have sufficient computing power. Neither the hardware nor the money to rent. (see #5)
2. The licensing implications of using the OpenAI API (self-instruct / Alpaca style training) are not entirely clear. (see #6)
3. The LLaMA training code is not open-sourced. (see #4)

Relevant Links

• LLaMA
  • Blog: Introducing LLaMA: A foundational, 65-billion-parameter large language model
  • arXiv: LLaMA: Open and Efficient Foundation Language Models
  • LLaMA Model Card
• GitHub: ChatLLaMA
• Alpaca
  • Blog: Alpaca: A Strong, Replicable Instruction-Following Model
  • GitHub: Stanford Alpaca: An Instruction-following LLaMA Model
  • arXiv: Self-Instruct: Aligning Language Model with Self Generated Instructions
• Training Data
  • Common Crawl
  • OSCAR
  • GC4 Corpus
  • Online Language Modelling Dataset Pipeline
  • Cleaned Alpaca Dataset

Contribution & Joining our Community

Our commitment to open source means that we are enabling - in fact encouraging - all interested parties to contribute and become part of our community. Contribution and feedback is encouraged and always welcome.

We communicate via a Slack channel. To get access to the Slack, please reach out to omar at huggingface.co or philipp at huggingface.co to be added.

Licensing

Copyright (c) 2023 by the LEL-A team

Licensed under the MIT License (the "License"); you may not use this file except in compliance with the License. You may obtain a copy of the License by reviewing the file LICENSE in the repository.