How does ReadyAI ensure annotation quality if miners are unsupervised?

Validators establish ground-truth tags by annotating full content with an LLM, then score each miner's window-level tags against those embeddings using cosine similarity. Miners who submit low-quality, off-topic, or duplicate tags are penalized and receive lower scores, which reduces their future task allocation and rewards. The scoring mechanism is transparent and on-chain, making it impossible for miners to game the system without genuinely improving accuracy.

Can I run ReadyAI on my own proprietary data instead of using ReadyAI's conversation server?

Yes. Validators can run custom conversation servers and point ReadyAI to any data source (customer databases, internal documents, proprietary datasets). The subnet is designed so each validator can sell access to their data-processing pipeline. You define the data, establish ground truth, and miners annotate it—without sharing raw data with the network.

What types of data can miners annotate?

ReadyAI supports conversations, webpages, surveys, documents, and skill specifications. For each type, miners generate semantic tags (topical labels), embeddings (vector representations), and metadata (named entities, summaries, classifications). The scoring mechanism adjusts based on task type; for example, named-entity extraction uses different penalty rules than conversation tagging because entity lists are naturally shorter and differently shaped than topic tags.

What are the compute requirements to run a miner or validator?

Miners and validators using OpenAI or Anthropic APIs need a CPU with at least 4 cores, 2.5 GHz per core (8 cores / 3.5 GHz recommended), 16 GB RAM, and 10 GB storage minimum (100 GB recommended). GPU-based miners require 8 GB VRAM minimum (24 GB recommended, NVIDIA A100 or equivalent). Validators using local LLMs require more GPU memory. All nodes need 100 Mbps download and 20 Mbps upload bandwidth.

How are miners paid for their work?

Miners earn TAO tokens from the Bittensor network based on their score. Scores are calculated each epoch (typically ~12 hours) and depend on annotation quality: higher cosine similarity to ground-truth embeddings, penalty avoidance, and volume of accepted tags all increase rewards. Top performers are allocated more future tasks. The incentive mechanism is designed so cost-effective miners (using cheap APIs or local models) can compete equally on quality with well-resourced operations.

Do I need to be a Bittensor expert to run a miner or validator?

No. ReadyAI provides Docker images, PM2 process management scripts, and Runpod deployment templates to simplify setup. You need a Bittensor wallet, an OpenAI (or alternative LLM provider) API key, and an internet connection. The README includes step-by-step guides for registration, environment configuration, and troubleshooting. Community support is available on Discord.

ReadyAI (SN33) — Distributed Data Tagging

Organizations need to convert raw unstructured data—conversations, documents, webpages, surveys—into structured, labeled datasets suitable for machine learning and vector databases. Manual human annotation is costly, slow, and difficult to scale; crowdsourcing introduces quality inconsistency. Current centralized services lock data into proprietary systems and require third-party API dependencies.

Decentralized validation networks can solve this by distributing annotation work across independent contributors who are economically incentivized to produce high-quality results. Instead of relying on a single company, validators establish a ground truth (the correct labels for full content), create windowed segments of data, and miners compete by analyzing those segments using LLMs. A transparent scoring mechanism—measuring tag embeddings against ground-truth embeddings using cosine similarity—rewards accuracy and penalizes low-effort submissions.

ReadyAI implements this on Bittensor (subnet 33). Validators pull raw data from any source (a conversation server, customer database, or proprietary dataset), tag the full content with an LLM to establish ground truth, split it into windows, and send masked windows to miners. Miners use configurable LLMs (GPT-4o, Claude, local models) to tag their assigned windows. Validators score each miner's tags against the ground truth, apply quality penalties, and the scoring directly adjusts miner incentives. The result: structured, tagged data at scale—and miners earn rewards based on quality, not volume.

Customer service teams can process transcripts of support conversations into structured, tagged datasets for training chatbots and improving response patterns, without hiring annotators.
Product researchers can analyze open-ended survey responses and customer feedback at scale, extracting consistent topic tags and sentiment markers across thousands of responses.
Knowledge base builders can convert unstructured documents, wikis, or internal procedures into semantically tagged, searchable repositories with embeddings for vector databases.
AI training teams can generate high-quality labeled datasets for fine-tuning domain-specific language models from raw text sources, avoiding crowdsourcing delays and inconsistency.
Content platforms can automatically extract metadata (topics, entities, summaries, skill tags) from user-generated content, webpages, or media transcripts for search, recommendation, and moderation.

ReadyAI is a decentralized data structuring and semantic tagging pipeline built on Bittensor that converts raw unstructured data into structured, AI-ready data. Validators establish ground truth by tagging full conversations/documents and create data windows for miners to process. Miners analyze these windows using LLMs to generate semantic tags and embeddings, which are scored against the validator's ground truth using cosine similarity metrics.

The subnet supports multiple task types including conversation tagging, webpage metadata generation, survey analysis, named entity extraction, and skill generation, enabling cost-efficient, scalable data annotation without human annotators.

The ReadyAI subnet operates on a validator-miner architecture where validators pull raw data (conversations, webpages, surveys, documents) from the ReadyAI Conversation Server, establish ground truth semantic tags via LLM annotation of full content, and create windowed data segments. Validators distribute masked task windows to miners, who use configurable LLMs (OpenAI GPT-4o default, also supports Anthropic, OpenRouter, Chutes) to generate semantic tags and embeddings. Miners return annotated results via the CgSynapse protocol.

Validators receive miner responses, score them using cosine similarity between miner tag embeddings and ground truth tag embedding neighborhoods, apply quality penalties, and update per-miner scores. These scores adjust miner weights in the metagraph, influencing future task allocation. Results are persisted to local SQLite databases and optionally logged to Weights & Biases for monitoring.

Customer service teams can process transcripts of support conversations into structured, tagged datasets for training chatbots and improving response patterns, without hiring annotators.
Product researchers can analyze open-ended survey responses and customer feedback at scale, extracting consistent topic tags and sentiment markers across thousands of responses.
Knowledge base builders can convert unstructured documents, wikis, or internal procedures into semantically tagged, searchable repositories with embeddings for vector databases.
AI training teams can generate high-quality labeled datasets for fine-tuning domain-specific language models from raw text sources, avoiding crowdsourcing delays and inconsistency.
Content platforms can automatically extract metadata (topics, entities, summaries, skill tags) from user-generated content, webpages, or media transcripts for search, recommendation, and moderation.

For tagging-style tasks (conversation_tagging, webpage_metadata_generation, survey_tagging, skill_generation): Each miner-submitted tag is scored by cosine distance from its embedding to the neighborhood of validator ground-truth tag embeddings. The final miner score blends: top-3 unique tag score mean (55% weight), overall mean tag score (25% weight), median tag score (10% weight), and single best tag score (10% weight). Penalties are applied for: not providing tags shared with ground truth, providing fewer than minimum unique tags, or having all tags below a quality threshold.

For skill_coverage_evaluation: Scoring blends Section Coverage (test suite addresses validator-defined sections) and Skill Coverage (test suite represents generated skill), with LLM-as-judge validation of assertion correctness; fabricated assertions score zero. For named_entities_extraction: Uses NoPenaltyGroundTruthTagSimilarityScoringMechanism skipping tag count penalties since entity extraction legitimately yields different result shapes. Scores update weights for miners in the metagraph, determining their proportion of future task allocation and earnings.

ReadyAI

Registration Activity

ReadyAI

Subnet Configuration

Market & Economics

For Miners

For Validators & Stakers

Registration Activity

What does ReadyAI do?

Use Cases

About

Subnet History

Frequently Asked Questions

Development Activity

Recent Activity

Hardware Requirements

What does ReadyAI do?

Use Cases

About

Subnet History

Frequently Asked Questions

Subnet Configuration

Market & Economics

For Miners

For Validators & Stakers

Hardware Requirements

Miner Setup

Validator Setup

Overview

Registration Activity

Participate

Overview

Subnet Configuration

Market & Economics

For Miners

For Validators & Stakers

Registration Activity

What does ReadyAI do?

Use Cases

About

Subnet History

Frequently Asked Questions

Participate

Development Activity

Recent Activity

Hardware Requirements

What does ReadyAI do?

Use Cases

About

Subnet History

Frequently Asked Questions

Subnet Configuration

Market & Economics

For Miners

For Validators & Stakers

Hardware Requirements

Miner Setup

Validator Setup