CamelGPT 🐪

A Proof of Concept of achieving high-performance generative text with minimal footprint and compute power.

Brought to you by the GPT-Research team

CamelGPT Demo

This demo showcases the power of CamelGPT, enabled by the Converters library. All computations and loading are handled client-side, reducing server load and maximizing efficiency.

Efficient Model Retrieval

Customization at Your Fingertips

Wide Range of Model Support

Optimized for Efficiency

Abstract

CamelGPT can generate text at blazing-fast speeds on low-end CPUs while outperforming prior models like GPT3 and Bert in instruction based tasks by a very significant margin. Unlike models that are limited to single-digit speeds with a lot of expensive compute, CamelGPT can achieve these high speeds with orders of magnitude less compute.

Approach

The CamelGPT architecture represents a significant breakthrough in the field, using both neural and linguistic features to generate coherent, grammatical text. The model combines recurrent and convolutional neural models and uses a new technique called “Eager Precached Dynamic Pruning” to reduce its compute requirements. We evaluate the model’s performance on the task answering benchmark Open LLM Leaderboard (our private version) but believe the results indicate its performance on other text generation tasks.

Model Details

In this paper, we trained an instruction-tuned model, CamelGPT-mini, as proof of concept of CamelGPT's new architecture
Model Date: May 2023
Model Type: Language
Dataset: Private
Parameters: 50k
Training: Training was performed in three phases. The first phase was an initial training run with only the convolutional component. The same initial network was trained with a linear embedding as in BERT in the second phase. Finally, the final architecture was introduced using the language models from the second and third phases and the final model.
Limitations: Our current model only supports english and may contain biases and produce inaccurate responses.

Demo

⚡ CamelGPT is so performant it can run on a browser via Converters

Inference

CamelGPT model weights are designed to be loaded and inferred using our Converters package, as seen in the demo below:


import { TextGenerationPipeline } from 'https://esm.sh/gh/gpt-research/converters/src/converters.js';

const main = async () => {
  // Initialize the pipeline with the desired model
  const pipeline = await TextGenerationPipeline("@gpt-research/CamelGPT-mini");

  // Generate text using the pipeline
  const generatedText = pipeline("Write a poem about camels.");

  // Log or use the generated text
  console.log(generatedText);
};

main();