██████╗ ██████╗ ███╗ ██╗██████╗ ███████╗███╗ ██╗███████╗███████╗ █████╗ ██╗
██╔════╝██╔═══██╗████╗ ██║██╔══██╗██╔════╝████╗ ██║██╔════╝██╔════╝ ██╔══██╗██║
██║ ██║ ██║██╔██╗ ██║██║ ██║█████╗ ██╔██╗ ██║███████╗█████╗ ███████║██║
██║ ██║ ██║██║╚██╗██║██║ ██║██╔══╝ ██║╚██╗██║╚════██║██╔══╝ ██╔══██║██║
╚██████╗╚██████╔╝██║ ╚████║██████╔╝███████╗██║ ╚████║███████║███████╗ ██║ ██║██║
╚═════╝ ╚═════╝ ╚═╝ ╚═══╝╚═════╝ ╚══════╝╚═╝ ╚═══╝╚══════╝╚══════╝ ╚═╝ ╚═╝╚═╝
| Component | Link |
|---|---|
| 🌐 Condense-AI & API Document | Visit Condense-AI |
| 📚 API Library | Explore API Library |
| 🔗 Organic Forwarder For Validators | Check Organic Forwarder |
| 📊 Miner Leaderboard & Statistics | View Miner Dashboard or Wandb Logger |
- (25/11/2024) Version 0.0.2 Update: Added condensing activations layers, Switched to Distributed Storage from Restful API Transfer, Emissions now allocated only to the top 30% miners.
- 🌐 Seamless Integration: Effortlessly integrates with LLM inference engines, such as transformers 🤗, vllm.
- 🧩 Token Compression: The subnet API compresses long sequences of natural language tokens into soft tokens.
- 🏛️ Decentralized Network: The subnet is a decentralized network that allows miners to contribute to the compression process.
- 📊 Tiered System: The subnet has a tiered system, with a research tier for experimentation and an inference tier for production-scale use. Incentive distribution is splitted for each tier.
- 📏 Benchmarking and Validation: The subnet owner defines synthetic metrics to benchmark miners’ performance, ensuring quality and efficiency.
| Tier | Purpose | Context Size | Incentive Percentage | Supporting Models |
|---|---|---|---|---|
research |
Optimize text-to-kv cache for a specific model | Up to 15000 characters | 60% |
mistralai/Mistral-7B-Instruct-v0.2 |
universal |
Compress text representation for various models | Up to 15000 characters | 40% |
meta-llama/Llama-3.1-8B-Instruct |
Supporting models can be flexibly added based on tailored need.
- Setup for miners: Miner Setup
- Setup for validators: Validator Setup
- Mechanism: Mechanism