We use a Large Language Model (LLM) trained on Llama 2C. We create and synthesize data with Python, clean the data, simulate a chess environment, train an AI model and a tokenizer. The model is compressed into a binary file and uploaded to a canister. This demonstrates the capabilities of a fully on-chain LLM in an Internet Computer Protocol (ICP) canister. The backend code is written in C++, and the model is uploaded as a list of bytes that is executed on the chain for inference.
The final product is a backend and frontend chess engine to play with the AI!. All full on chain and inference!.
OISY-WALLET: dm5n3-2x7rv-3fbed-or75y-a4lwl-xlpt6-memhv-jefhe-u2xfp-2gvw4-3qe
REPO-URL: https://github.com/bridge-23/chess23
CANISTER FRONTEND:https://cirh4-aiaaa-aaaan-qmvqa-cai.icp0.io/
CANISTER BACKEND: https://a4gq6-oaaaa-aaaab-qaa4q-cai.raw.icp0.io/?id=huk5z-laaaa-aaaan-qmvpq-cai
What we learned: New capabilities of specialized models in ICP with large languaje models , this model even when small can be really powerfull if specialized in small tasks.
We are proud of deploying the first chess LLM model on the mainnet that can fully understand the chess board and make moves to play against a player, the model is only 25mb and was trained for 2 hours to accomplish this.
laughtt
commented
2 months ago
We use a Large Language Model (LLM) trained on Llama 2C. We create and synthesize data with Python, clean the data, simulate a chess environment, train an AI model and a tokenizer. The model is compressed into a binary file and uploaded to a canister. This demonstrates the capabilities of a fully on-chain LLM in an Internet Computer Protocol (ICP) canister. The backend code is written in C++, and the model is uploaded as a list of bytes that is executed on the chain for inference. The final product is a backend and frontend chess engine to play with the AI!. All full on chain and inference!. OISY-WALLET: dm5n3-2x7rv-3fbed-or75y-a4lwl-xlpt6-memhv-jefhe-u2xfp-2gvw4-3qe REPO-URL: https://github.com/bridge-23/chess23 CANISTER FRONTEND:https://cirh4-aiaaa-aaaan-qmvqa-cai.icp0.io/ CANISTER BACKEND: https://a4gq6-oaaaa-aaaab-qaa4q-cai.raw.icp0.io/?id=huk5z-laaaa-aaaan-qmvpq-cai What we learned: New capabilities of specialized models in ICP with large languaje models , this model even when small can be really powerfull if specialized in small tasks. We are proud of deploying the first chess LLM model on the mainnet that can fully understand the chess board and make moves to play against a player, the model is only 25mb and was trained for 2 hours to accomplish this.