MTSAIR/Cotype-Nano

MTSAIR/Cotype-Nano: A Lightweight, Efficient LLM

MTSAIR/Cotype-Nano is a 1.5 billion parameter large language model specifically engineered for resource-constrained environments. Its primary design goal is to deliver fast and efficient interaction while maintaining high performance, making it ideal for applications where computational resources are limited.

Key Capabilities & Training

• Optimized for Efficiency: Designed to operate effectively with minimal resources, ensuring quick response times.
• Two-Stage Training: The model underwent a two-stage training process:
  • Stage 1: MLP layers were trained on a combination of mathematics and code datasets.
  • Stage 2: The entire model was further trained on internal and open synthetic instructional datasets, enhancing its ability to follow instructions.
• Multilingual Support: The README provides examples and system prompts in Russian, indicating potential multilingual capabilities, particularly for Russian language tasks.

Performance Highlights

• ru-llm-arena Score: Achieved a score of 30.2 (local measurement) on the ru-llm-arena benchmark, outperforming several other models in its class, including vikhr-it-5.3-fp16-32k and Qwen-Qwen2.5-1.5B-Instruct.
• 4-bit Quantization: A 4-bit quantized version, Cotype-Nano-4bit, also performed competitively with a score of 22.5.

Good For

• Edge Devices & Low-Resource Deployments: Its lightweight nature makes it suitable for deployment on devices with limited memory and processing power.
• Applications Requiring Fast Inference: Ideal for scenarios where quick, real-time responses are crucial.
• Instruction Following: Benefits from training on synthetic instructional datasets, making it capable of generating detailed and expanded answers to user queries.
• Code Generation & Mathematical Tasks: Initial training on mathematics and code suggests proficiency in these domains.