Mxode/NanoLM-365M-Base
TEXT GENERATIONPricing:Input $0.04 / Cached $0.008 / Output $0.08Concurrent Unit Cost:1Model Size:0.5BQuant:BF16Context Size:32kTool Calling:SupportedPublished:Sep 8, 2024License:gpl-3.0Architecture:Transformer Open Weights Featherless Exclusive Warm
Mxode/NanoLM-365M-Base is a 365 million parameter language model developed by Mxode, based on Qwen2-0.5B. It features a reduced parameter count achieved by replacing the original tokenizer with BilingualTokenizer-8K. This model is specifically optimized for efficient deployment and fine-tuning on downstream tasks, making it suitable for applications requiring a compact yet capable language model.
Loading preview...
NanoLM-365M-Base Overview
NanoLM-365M-Base is a compact language model developed by Mxode, derived from the Qwen2-0.5B architecture. Its primary innovation lies in the replacement of the original tokenizer with the more efficient BilingualTokenizer-8K, which successfully reduced the total parameter count from 0.5 billion to 365 million.
Key Characteristics
- Parameter Efficiency: Achieves a significantly smaller footprint (365M parameters) compared to its base model, Qwen2-0.5B, through tokenizer optimization.
- Targeted Training: The model's embedding layer was specifically trained for 40,000 steps on a combination of Chinese and English datasets, including wikipedia-zh and cosmopedia-100k, while freezing the backbone parameters.
- Fine-tuning Readiness: This training approach aims to recover performance and facilitate easier fine-tuning for various downstream applications.
Ideal Use Cases
- Resource-Constrained Environments: Excellent for deployment where computational resources or memory are limited.
- Downstream Task Adaptation: Well-suited as a base model for further fine-tuning on specific tasks, benefiting from its optimized embedding layer.
- Research and Development: Provides a lightweight yet capable foundation for experimenting with language model applications.