TinyChat15M is a 15-million parameter conversational language model built on the Meta Llama 2 architecture. Designed to operate on devices with as little as 60 MB of free memory, TinyChat15M has been successfully deployed on the Sipeed LicheeRV Nano W, a compact RISC-V development board equipped with just 256 MB of DDR3 memory. Inspired by Dr. Andrej Karpathy’s llama2.c project, TinyChat15M showcases that small conversational language models can be both effective and resource-efficient, making advanced AI capabilities more accessible and sustainable. You can find a detailed blog post on this project here.
First, navigate to the folder where you keep your projects, and then clone this repository into that folder:
git clone https://github.com/starhopp3r/TinyChat.git
Next, navigate to the llama2.c folder:
cd llama2.c
Now, download the TinyChat15M model from Hugging Face:
wget https://huggingface.co/starhopp3r/TinyChat/resolve/main/TinyChat15M.bin
Next, compile the C code:
make run
Now, to run the TinyChat15M Assistant use the following command:
./run TinyChat15M.bin -t 1.0 -p 0.9 -n 2048 -m chat
Note that the temperature (-t flag) and top-p value (-p flag) can be set to any number between 0 and 1. For optimal results, it's recommended to sample with -t 1.0 and -p 0.9, meaning a temperature of 1.0 (default) and top-p sampling at 0.9 (default). Intuitively, top-p sampling prevents tokens with extremely low probabilities from being selected, reducing the chances of getting "unlucky" during sampling and decreasing the likelihood of generating off-topic content. Generally, to control the diversity of samples, you can either adjust the temperature (i.e., vary -t between 0 and 1 while keeping top-p off with -p 0) or the top-p value (i.e., vary -p between 0 and 1 while keeping the temperature at 1), but it’s advisable not to modify both simultaneously. Detailed explanations of LLM sampling strategies can be found here, here and here.