Back to blog

Mastering LLM Training with Appy Pie: Embracing the No-Code Revolution

---

Overview

Large Language Models, such as OpenAI’s GPT-4 or Google’s PaLM, have revolutionized the AI landscape. However, many organizations lack the resources to train these models independently and rely on a select few tech giants for their AI solutions.

At Appy Pie, we have dedicated substantial resources to build the infrastructure necessary for training our Large Language Models from scratch, concentrating on no-code development. In this article, we will outline our LLM training process, from raw data handling to deployment in user-oriented production environments. We will delve into the engineering challenges we face and how we utilize the services of modern LLM vendors like Dolly by Databricks, StableLM Alpha 7b by Stability AI, Hugging Face, and MosaicML.

Although our models are mainly designed for no-code development applications, the techniques and insights shared here are relevant to all LLMs, including general-purpose language models. We will provide in-depth information about our methods in a series of blog posts in the coming weeks.

The Benefits of Training Your Own LLMs

Appy Pie focuses on customization, independence, and cost-effectiveness when training our LLMs.

1. Customization: By training a tailored model, we can address specific requirements such as platform-exclusive features, terminology, and context not covered well by general-purpose models like GPT-4 or code-focused models like Codex.
2. Independence: Reducing dependency on a few AI providers is beneficial not only for Appy Pie but also for the larger developer community. By training our models, we can open-source some of them, which would be impossible without the capacity to train them.
3. Cost-effectiveness: LLMs are often too expensive for widespread adoption. Appy Pie aims to make cutting-edge AI accessible to all by training custom models that are smaller, more efficient, and can be hosted at significantly lower costs.

Data Pipelines

Establishing efficient data pipelines is crucial for LLM training, given the vast amount of data required. Our pipelines are highly optimized and flexible, making it easy to incorporate new public and proprietary data sources.

Utilizing The Stack

We begin with The Stack dataset from Hugging Face, which contains approximately 2.7 TB of permissively licensed source code in over 350 programming languages. Hugging Face offers several helpful tools, such as tokenization, model inference, and code evaluation.

Data Processing with Databricks

We employ Databricks for advanced data processing and pipeline construction. This approach also facilitates the integration of additional data sources in the future.

Tokenization and Vocabulary Training

We train custom vocabularies on random subsamples of our training data, enhancing model performance and expediting model training and inference.

Model Training with MosaicML

We train our models using MosaicML, which provides numerous advantages, such as access to GPUs from various cloud providers, well-optimized LLM training configurations, and managed infrastructure for orchestration, efficiency optimizations, and fault tolerance.

Evaluation

We utilize the HumanEval framework to test our models. We generate a code block for a specific function signature and docstring, and then run test cases on the generated function to assess if the code block functions as intended.

Production Deployment

We deploy our trained and evaluated models in production using NVIDIA’s FasterTransformer and Triton Server for swift, distributed large model inference. We can also autoscale our models to accommodate demand through our containers infrastructure.

Feedback and Iteration

At Appy Pie, we recognize that the key to continuous improvement lies in the ability to gather feedback and iterate on our models. Our no-code-focused model training platform is designed with rapid iteration in mind, allowing us to efficiently update and refine our models based on user feedback and changing requirements.

By deploying our models in production, we can actively collect user feedback and identify areas for improvement. This real-world data is invaluable in helping us fine-tune our models and ensure they remain relevant and effective for our user base.

Our platform is also built to be adaptable, capable of accommodating changes in data sources, model training objectives, and server architecture. This flexibility allows us to keep pace with the ever-evolving world of AI and stay at the forefront of new developments and capabilities.

In the future, we plan to further enhance our platform by integrating Appy Pie itself into the model improvement process. This will involve leveraging techniques such as Reinforcement Learning Based on Human Feedback (RLHF) and instruction-tuning using data collected from Appy Pie user interactions. This will enable us to harness the power of our no-code development expertise to create an even more powerful and versatile AI-driven experience for our users.

Thanks for reading our post. If you’re curious about the latest language models, be sure to check out our blog on StableLM Alpha 7b vs Dolly and The Expanding World of Large Language Models.