- Quickstarter
- Introduction
- Installation
- Writing your own incentive mechanism
- Writing your own subnet API
- Subnet Links
- License
// TODO // * Validators: link // * Miners: link
Crynux, derived from "crystallization" and "nucleation" of Decentralized AI (#DeAI), symbolizes the organic growth and structure formation akin to crystal formation. Our mission is to establish a robust DeAI framework, embracing the decentralized ethos.
Crynux means crystallization and nucleation of Decentralized AI (#DeAI), symbolizes the organic growth and structure formation of crystal. Our mission is to build a 100% DeAI framework.
Crynux has launched:
- Hydrogen Network: Decentralized inference
- Helium Network: Dynamic task schedule
- Now working on Lithium Network for decentralized finetune. This is just the beginning. Crynux remains dedicated to the ongoing evolution of the network, mirroring nuclear fusion.
Crynux Orbital is a BitTensor subnet for decentralized retrieval augmented learning. BitTensor's on-chain consensus engine Yuma Consensus provides the essential incentive mechanism for DeAI model development, emphasizing data integrity, alignment, and evaluation for optimal service quality.
Decentralized incentive mechanism represents a paradigm shift towards a more democratic, trustworthy, and superior approach to AI model training and evaluation.
DeRAG stands as a groundbreaking technology aimed at enhancing Large Language Models (LLMs), rendering them more factual, stable, and reliable. LLMs have garnered notoriety for their tendency towards hallucinatory outputs, potentially propagating misinformation. Leveraging decentralized efforts, DeRAG strategically retrieves pertinent information to augment and validate text generated by LLMs.
Google's misstep during the Bard demo, intended as a response to OpenAI's ChatGPT, proved costly, resulting in a staggering $100 billion loss in market value. Read the news here.
In this demo, a user asked: "What new discoveries from the James Webb Space Telescope can I tell my 9 year old about?"
In response, Bard erroneously stated, "JWST took the very first pictures of a planet outside of our own solar system."
However, according to NASA, "the first image showing an exoplanet was taken by the European Southern Observatory’s Very Large Telescope nearly".
The solution is already demonstrated in the example: use DeRAG.
- LLM generates some text, as in the example: "JWST took the very first pictures of a planet outside of our own solar system"
- Validators issue requests to miners.
- Miners furnish evidence either supporting or contradicting the statement. For instance, in the example, miners provide information sourced from NASA or other relevant resources indicating that the European Southern Observatory captured the first image of an exoplanet.
- Validators scrutinize the outcomes and assign reward weights to miners based on their contributions.
Our team member led the effort to employ RAG in Google's products. Read the news here. Now we build the decentralized enhancements to democratize and fortify the entire process, ensuring greater transparency and trustworthiness.
Empowered by Crynux's decentralized computing layer, validators issue or relay LLM prompts and outputs to the network, expecting Miners to provide relevant references.
Each result provided by miners undergoes evaluation based on the following criteria:
- Reliability: Miners are expected to offer authentic and trustworthy information, devoid of falsehoods.
- Precision: Miners must accurately provide pertinent references, avoiding arbitrary guesses.
- Recall: Miners should present comprehensive references, ensuring no critical details are missed.
The reward score will be calculated as:
2* Reliability* (Precision * Recall) / (Precision + Recall)
Utilizing Crynux decentralized computing layer, miners engage with validators by furnishing references that either substantiate or challenge the given statement. The process involves the following steps:
- Preprocess: Decontexting the text, segmenting it into multiple individual statements.
- Query Generation: For each statement, generating search queries to extract relevant information.
- Information Retrieval: Conducting searches across various sources to obtain pertinent data corresponding to each query.
- Filtering and scoring: Evaluating each search result, filtering out irrelevant information, and assigning a score to each statement based on its relevance and credibility.
We will build Discord bot that equips with Crynux Orbital in our discord group, and provide a mechanism for community members to invovle into this mechanism.