Learn more at Hugging Face or visit the Arynwood Terminal.
This project started with a simple goal, to understand what makes a chatbot feel alive. Using the free tier on Hugging Face, I worked with the zephyr-7b-beta model, a base that was already lightweight enough to experiment with but still powerful enough to hold a real conversation.
Because of hardware constraints, I did not train a new model from scratch. Instead, I focused on direct system message editing and adjusting runtime parameters like temperature, top-p sampling, and token limits. Temperature controls how much randomness the bot uses. Higher values make it more creative and unpredictable, lower values make it more logical and focused. Top-p sampling changes how many possible next words the bot can choose from, balancing between creativity and precision. Max new tokens determines how much text the model is allowed to generate in a single response, which also helps prevent it from running out of memory on small servers.
Inside the app.py file, I built a custom interface with sliders for these controls, so every conversation could be shaped in real time without needing to reprogram the bot. By shaping the system message carefully, I could tune the bot’s tone, style, and internal assumptions without touching the model weights. This gave me surprising freedom to craft a personality that fits the atmosphere I want for Arynwood, without any heavy compute resources or paid infrastructure.
While zephyr-7b-beta cannot match the depth of commercial models like GPT-4, it does something just as important. It gives me control. It lets me see what happens when you adjust the temperature to 1.2 instead of 0.7, or when you open up the top-p to allow more creative drift.
The hands-on experience has been essential. It has shown me how lightweight language models can be customized, how system messages act like the skeleton of a bot’s mind, and how important it is to work within memory and inference limits when you are not relying on expensive APIs.
This version of Veylan is a starting point. I plan to expand this work by hosting multiple small bots locally on my own machine, each one adapted for different tasks like coding, creative writing, and self-guided learning. But everything begins here, with a simple Hugging Face deployment, a modified system prompt, and a desire to understand the architecture of intelligence.