Subnet 5 welcomed a major update at the end of September, and in this article we will introduce its specific content to help the market re-understand and evaluate the importance and potential of Subnet 5.
Subnet 5 Open Kaito
Emission: 0.88% (2024–10–10)
Github: https://github.com/OpenKaito/openkaito
What is SN5?
The core goal of Bittensor Subnet 5 is to develop the world's best-performing and most widely-applicable text embedding model, and to realize decentralized model training, evaluation, and service, while opening up these models in the form of APIs to support a wide range of downstream applications.
SN 5 utilizes the Bittensor decentralized node network to achieve dynamic evaluation and continuous improvement of the models.
The Miners of SN 5 are responsible for training models using continuously updated large-scale text corpora, and committing to provide the models with low latency and high throughput to respond to requests from downstream applications.
The Validators of SN 5 are responsible for using a series of benchmarks to rigorously evaluate the models, and continuously screen out higher-quality models. The models provided by the Miners will be compared with the best text embedding models on the market to ensure that the models produced by SN 5 are always competitive and continuously evolving.
How to evaluate Miners?
In SN 5, Miners will receive batches of text and use the models they provide to perform text embedding, while Validators will evaluate the embedding quality by comparing the learning loss function.
The formula of the loss function is:
Where c is the target embedding, x is the positive sample, and x' is the negative sample. This process aims to maximize the mutual information between the positive sample x and the target embedding c.
Miners with higher embedding quality will receive higher scores, and the models they provide will be considered higher quality.
What kind of output has SN 5's update achieved?
Under this incentive mechanism, the models of SN 5 products have already achieved certain results:
On October 9, 2024, the text embedding model provided by the Miners achieved improvement compared to the benchmark of Openai.
As the quality of the models produced by SN 5 improves, users will be able to access general text embedding models that surpass the performance of existing state-of-the-art models. These models will be publicly available through the Validator API of the Bittensor Subnet 5, making it convenient for widespread application and integration into various applications.
How to evaluate the brand new SN 5?
SN5 optimizes and trains text embedding models in a decentralized way, and can make a huge contribution to the Bittensor ecosystem and decentralized AI:
- Promote the evolution of the most widely-applicable models in the industry: The goal of SN 5 is to develop the world's best-performing and most widely-applicable text embedding model. The models produced by SN 5 will be evaluated against an infinitely large and dynamic dataset to ensure the highest level of domain generalization capability.
- Fully utilize the decentralized network to ensure dynamic evaluation and continuous improvement: Utilizing the Bittensor network for decentralized training and model improvement avoids the limitations of centralized AI model development, and enhances the transparency and censorship resistance of the system. And through the dynamic evaluation of Validators, it continuously incentivizes Miners to optimize model performance, ensuring that the embedding models of SN5 not only surpass the existing state-of-the-art models, but also adapt to the latest real-world knowledge, and remain competitive and push the boundaries of industry performance.
