Chaturanga Mini
Chaturanga is a lightweight SLM designed for playing chess using natural language move generation. It is based on the Supra family (https://huggingface.co/collections/SupraLabs/supra2) which allows it to generate (mostly)accurate and legal moves whilst keeping memory footprint to a minimum. It's small size (25M-100M parameters) allows it to run comfortably on laptop grade GPUs and CPUs.
Chaturanga Mini is a 50M variant based on the Supra-50M-Base which has been finetuned on 400,000 rows from the chessllm dataset.
The Mini variant is nearly double the size of the Nano variant (25M -> 50M), which allows it to generate more comprehensive and non-illogical moves whilst keeping similar inference times. Despite its small size, it can still generate fairly precise and legal moves in SAN.
The Chaturanga model comes in the following sizes: Nano, Neo(Nano+), Mini, Pro and Ultra(Pro+)
Estimated parameters: ~50M
Architecture: Llama
Intended use: as an experimental chess-playing SLM
Training data
- Source: chessllm dataset (https://huggingface.co/datasets/mlabonne/chessllm)
- Rows: ~400,000 rows templated in SAN (Standard Algebraic Notation)
- Training: trained for 7,000 steps on an RTX 3050 (4GB VRAM)
Usage
The model and the chess GUI can be found in the chaturanga_gui.py file included in the uploads.
Limitations
- The model only has 50M parameters, so hallucinations and illegal moves may occur.
- Model for experimental use only; users should employ it as such under license.