Listen to this story
|
Meta’s AI has a reputation for their willingness to open source their models to the academic community. Their latest model, LLaMA, is no exception, with the model’s weights available to interested researchers on a case-by-case basis.
However, shortly after its release, the model and its weights were leaked and made available for download through torrents. This breach prompted GitHub user Christopher King to submit a pull request to the LLaMA GitHub page, which included a link to the open model via torrent.
Despite the unauthorised release, the developer community has taken full advantage of the newfound availability. They have optimised the model to operate on even the most basic of devices, introduced additional functionality, and even employed LLaMA to create use cases.
Here are 7 ways in which the LLaMA model has been used by the community since its release.
Stanford Alpaca
Stanford University researchers developed a model called ‘Alpaca’, which is a fine-tuned version of ‘LLaMA 7B’. Using more than 50,000 demonstrations that follow instructions from GPT 3.5, the researchers trained Alpaca to produce outputs that are comparable to those generated by OpenAI’s model. Remarkably, the training and inferencing costs of this model were a mere $600, a significant reduction from the millions of dollars usually required to train these models.
Upon the model’s open-sourcing, researchers recognised Alpaca’s potential. While the initial applications were modest, such as creating a Homer Simpson bot, the model quickly demonstrated a broad range of valuable uses.
Check out the GitHub repository here.
GPT4ALL
Nomic AI entered the race to compete with other companies that are exploring GPT models with ‘GPT4All’. This language model, with 7B parameters, was trained on a carefully curated corpus of over 800,000 high-quality assistant interactions gathered using the GPT-Turbo-3.5 model.
Inspired by Stanford’s ‘Alpaca’, GPT4All has produced approximately 430,000 high-quality assistant-style interaction pairs, including story descriptions, dialogue, and code. The researchers behind the model used numerous instances of Meta’s LLaMA language model for training.
Currently, GPT4All is only licensed for research purposes, as it is based on Meta’s LLaMA, which has a non-commercial licence. However, it also comes in a quantised 4-bit version, allowing users to run the model on consumer-grade hardware with limited computational resources by accepting less precision during training.
Check out the GitHub repository here.
Vicuna-13B
Researchers from UC Berkeley and more recently introduced a new open-source alternative to GPT-4 named ‘Vicuna-13B‘. The model boasts impressive results, achieving 90% of ChatGPT’s quality, while the training cost is merely $300. This was made possible through fine-tuning the model with LLaMA and incorporating user-shared conversations gathered from ShareGPT.
The contributors of Vicuna-13B have emphasised the model’s superior natural language processing capabilities when compared to other models, including ChatGPT. While there are some similarities between the two models, Vicuna sets itself apart through its efficiency and customisation features.
Check out the GitHub repository here.
Koala
UC Berkeley unveiled ‘Koala’, a new dialogue model for research purposes. The model has been trained using Meta’s LLaMA, fine-tuned with a high-quality dataset of dialogue data scraped from the web, with a particular emphasis on responses to queries from other large language models, such as ChatGPT.
To achieve the best possible dataset, the creators of Koala prioritised quality over quantity while scraping the web. In total, 60,000 dialogues publicly shared on ShareGPT were collected via APIs for the model’s training. However, the team eliminated redundant and non-English dialogues, ultimately shrinking the dataset down to around 30,000 dialogues.
Check out the GitHub repository here.
Guanaco
‘Guanaco‘ is an instruction-following language model, trained on Meta’s LLaMA 7B model with an additional 534,530 entries covering various linguistic and grammatical tasks in seven languages.
The developers stated that the model shows great promise in a multilingual environment after being optimised and retrained with this data. However, it has not been screened for harmful or explicit content, and developers must be cautious while using it for research or practical purposes.
Check out the repository here.
Dalai
Language modelling appears to be reaching the ‘Stable Diffusion’ stage, as evidenced by the launch of ‘Dalai‘, a user-friendly way to run LLaMa on local machines. The tool is capable of running on as little as 4 GB of RAM and does not require an internet connection, offering users complete control over their language model.
Dalai requires Python <= 3.10 and Node.js >= 18. For optimal performance, users may need to update their Node.js version, which can be done via a PPA on Ubuntu 22.04 LTS test systems.
Check out the GitHub repository here.
Simple WebUI
A tool called ‘Simple LLaMA Finetuner’ has been introduced to help beginners fine-tune the LLaMA-7B language model using the LoRA method with the PEFT library on standard NVIDIA GPUs.
The tool is designed with an intuitive interface and can run on a regular Colab Tesla T4 instance with small datasets and sample lengths of 256. The tool also enables users to efficiently manage their dataset, customise parameters, train the model, and evaluate its inference capabilities.