Apply for a GPU community grant: Personal project

#1
by PlusBIntelligenceLab - opened

Treevik-2 is a 2.5 Billion parameter, 8-bit quantized, instruction-tuned language model developed from scratch by PlusB Intelligence Lab (an independent research initiative by Sayan Chowdhury based in India). My primary goal with Treevik is to experiment with and optimize medium-scale LLM architectures to push the boundaries of what independent researchers can achieve.

Currently, I am personally funding the compute required to train these models out of my own pocket. My focus is on allocating my limited personal budget towards the research and the actual training compute. However, to iterate and create better models, I need continuous feedback from the community.

I’ve deployed Treevik-2 on a Gradio ChatInterface on Hugging Face Spaces for public testing. Unfortunately, the free basic CPU tier makes inference painfully slow (taking multiple seconds per token) for a 2.5B parameter model, which discourages users from interacting with it and slows down my feedback loop significantly.

How the GPU will be used: The community GPU grant will be used exclusively to power the inference environment for the Treevik-2 Gradio Space. Having access to a GPU (like a T4 or L4) will drastically speed up response times for the public chat interface, enabling a smooth, realtime user experience. This will allow me to gather more usage data, collect high-quality feedback on model behaviors and edge cases, and directly inform the architectural improvements for the upcoming Treevik iterations, while I continue to fund the heavy training loops myself.

Sign up or log in to comment