Loading
ItsAI is a bittensor subnet focused on high-quality AI detection for texts. Recognised as the most accurate AI detector by MGTD benchmark.
Neuron registration trends
Organizations and educators increasingly struggle to distinguish human-written content from text generated by large language models. Manual review is time-consuming and unreliable, while existing detection tools often fail on diverse writing styles, languages, and generation parameters. A centralized approach creates a single point of failure and limits adaptability as new models emerge.
A decentralized network of detection nodes addresses this by distributing the work of building and improving detection models across many contributors, each incentivized to achieve higher accuracy. Contributors can experiment with different detection approaches, data sources, and model architectures. Those whose detectors perform best on a continuous stream of validation challenges receive rewards proportional to their accuracy, creating ongoing competition that drives innovation without relying on any single team.
Subnet 32 implements this as a distributed detection system where miners run inference models (based on perplexity analysis or transformer classifiers) on text samples provided by validators. Validators continuously generate test data from human sources and AI models across 30+ language models with varied generation parameters, then measure each miner's precision, recall, and ranking quality. The top performers accumulate stake and influence, while the network's collective detection capability remains current with evolving LLM capabilities.
Use cases were auto-generated from the GitHub repository and may not reflect the latest changes.
Subnet 32 is a decentralized AI-generated content detection system that identifies whether text was written by humans or AI models. It leverages perplexity-based models and machine learning classifiers to provide both binary classification and sentence-level probability scoring, achieving state-of-the-art accuracy (92% ROC-AUC on MGTD benchmark) and powering commercial applications including a website, browser extension, API, and X/Twitter bot.
Miners receive TextSynapse requests containing lists of texts from validators. Each miner runs a pretrained model (either PPL-based using Phi-2 or DeBERTa classifier) to generate per-token predictions, encrypts predictions with validator's ephemeral public key to maintain privacy on-wire, and returns encrypted predictions. Validators generate test data by sampling from Pile dataset and generating AI completions using Ollama-served LLMs with random generation parameters, apply augmentations, and send batches to miners.
Validators decrypt responses and compare against ground truth labels, computing F1/FP/AP scores. Scores are aggregated in a moving average and used to set weights at epoch end. Metagraph tracks stake, validator status, and uid mappings for blacklisting and priority functions.
Validators continuously generate test data by prompting 30+ open-source SOTA LLMs (Llama, Qwen, Command-R, Mistral, Gemma, etc.) with prompts derived from the Pile dataset, then compare miner predictions against ground truth labels. The dataset includes data augmentation (misspellings, adjective removal) to prevent memorization. Miners receive requests with batches of texts and must return per-token predictions.
This information was auto-extracted from the GitHub repository and may be incomplete or out of date. Always refer to the subnet's GitHub repository for the latest instructions.
Scoring uses: (1) F1 score for classification accuracy, (2) FP_score = 1 - FP/total_samples to penalize false positives (human text misclassified as AI), (3) AP score for ranking quality across probability thresholds. Final reward is the average of these three metrics. Validators use softmax normalization on scores for miner motivation.
Out-of-domain F1 minimum threshold (0.9) ensures baseline quality.
The sections above were auto-generated from the GitHub repository and may not reflect the latest changes. View repository
ItsAI is a bittensor subnet focused on high-quality AI detection for texts. Recognised as the most accurate AI detector by MGTD benchmark.
ItsAI is a bittensor subnet focused on high-quality AI detection for texts. Recognised as the most accurate AI detector by MGTD benchmark.
FAQs were auto-generated from the GitHub repository and may not reflect the latest changes.
Holder & total-staked metrics as of 1h ago (hourly snapshot)
NVIDIA RTX A4000 with minimum 16 GB VRAM mentioned in mining docs as alternative. CUDA compute capability 6.0 minimum, 7.0 recommended. x86_64 architecture, DDR4 RAM, SSD storage with 1000 IOPS minimum.
CUDA compute capability 6.0 minimum, 7.0 recommended. x86_64 architecture, DDR4 RAM, SSD storage with 1000 IOPS minimum. Server provider must support systemd for ollama service auto-restart.
Update system and clone the repository
apt update && apt upgrade -y && git clone https://github.com/It-s-AI/llm-detection
Create and activate Python virtual environment
python -m venv venv && source venv/bin/activate
Install detection package and dependencies
cd llm-detection && rm -rf detection.egg-info && python -m pip install -e .
Verify installation of bittensor and detection packages
python -c "import bittensor, detection; print('ok', bittensor.__version__)"Download pre-trained DeBERTa model weights
wget https://huggingface.co/sergak0/sn32/resolve/main/deberta-large-ls03-ctx1024.pth -O models/deberta-large-ls03-ctx1024.pth && wget https://huggingface.co/sergak0/sn32/resolve/main/deberta-v3-large-hf-weights.zip -O models/deberta-v3-large-hf-weights.zip && unzip models/deberta-v3-large-hf-weights.zip -d models/deberta-v3-large-hf-weights
Create Bittensor wallet and hotkey
btcli w new_coldkey && btcli w new_hotkey
Register hotkey on subnet 32
btcli s register --netuid 32 --wallet.name YOUR_COLDKEY --wallet.hotkey YOUR_HOTKEY
Install PM2 process manager
sudo apt update && sudo apt install jq npm && sudo npm install pm2 -g && pm2 update
Start miner with PM2
pm2 start --name net32-miner --interpreter python3 ./neurons/miner.py -- --wallet.name YOUR_COLDKEY --wallet.hotkey YOUR_HOTKEY --neuron.device cuda:0 --axon.port 70000
Update system and clone the repository
apt update && apt upgrade -y && git clone https://github.com/It-s-AI/llm-detection
Create and activate Python virtual environment
python -m venv venv && source venv/bin/activate
Install detection package with custom mathgenerator
cd llm-detection && python3 -m pip install -e . && python3 -m pip uninstall mathgenerator -y && python3 -m pip install git+https://github.com/synapse-alpha/mathgenerator.git
Create Bittensor wallet and hotkey
btcli w new_coldkey && btcli w new_hotkey
Register hotkey on subnet 32
btcli s register --netuid 32 --wallet.name YOUR_COLDKEY --wallet.hotkey YOUR_HOTKEY
Set Weights & Biases API key for logging
export WANDB_API_KEY="<YOUR_WANDB_API_KEY>"
Install PM2 and hardware detection tools
sudo apt update && sudo apt install jq npm lshw && sudo npm install pm2 -g && pm2 update
Install Ollama for running LLMs
curl -fsSL https://ollama.com/install.sh | sh
Start Ollama service with PM2
pm2 start --name ollama "ollama serve"
Install build tools for cc_net compilation
sudo apt-get install build-essential cmake libboost-system-dev libboost-thread-dev libboost-program-options-dev libboost-test-dev zlib1g-dev libbz2-dev liblzma-dev zip unzip -y
Build and install cc_net for text quality filtering
cd cc_net && export TAR_OPTIONS="--no-same-owner" && pip install -e . && make install && make lang=en dl_lm && cd ..
Make run script executable
chmod +x run.sh
Start validator with PM2
pm2 start run.sh --name llm_detection_validators_autoupdate -- --wallet.name YOUR_COLDKEY --wallet.hotkey YOUR_HOTKEY --axon.port 70000 --neuron.device cuda:0