Overview

GroundLayer (Subnet 20) is a specialized Bittensor subnet that orchestrates a decentralized, competition-driven environment for evaluating and rewarding fine-tuned tool-calling language models (LLMs). Originally launched as BitAgent and later rebranded to Bounty Hunter, SN20 now operates under the GroundLayer banner, integrating on-chain governance and economic incentives to solve the core problem of trustless, transparent AI model benchmarking and rewarding within the Bittensor ecosystem.

Subnet Architecture and Consensus

As a first-class Bittensor subnet, GroundLayer adheres to the network’s Yuma Consensus interface, engaging miners and validators in a cyclical loop. Miners register their wallets and submit fine-tuned LLMs as on-chain neurons, while validators independently instantiate each submission in an offline environment to execute standardized task evaluations. Upon completion, validators publish performance metrics to the metagraph, enabling on-chain scoring and reward issuance.

Miner Operations

Participants wishing to mine SN20 begin by fine-tuning an 8-billion-parameter (or smaller) LLM for tool-calling tasks, often leveraging datasets from the Berkeley Function Calling Leaderboard (BFCL). After cloning the bitagent_subnet repository and installing dependencies, a miner registers via the Bittensor CLI: “btcli subnet register –subtensor.network finney –netuid 20”. Once live, miners submit their model names and await validator-driven BFCL evaluations in both single-turn and multi-turn modes.

Validator Responsibilities

Validators in SN20 must provision high-VRAM servers (minimum 48 GB per GPU) and install the BFCL evaluation suite. Using PM2 and vllm, validators automatically fetch submitted models from Hugging Face, spawn isolated evaluation servers, and run offline task batches. They record scores reflecting accuracy and generality, then aggregate these into moving averages on the Bittensor blockchain. Validator performance underpins the network’s trustworthiness and ensures unbiased reward distribution.

Output and End Product

The end product of SN20 is a continuously updated, on-chain leaderboard of tool-calling LLM submissions, accompanied by tokenized rewards distributed as dTAO (subnet tokens). Miners’ scores translate directly into emissions, fostering a transparent market where high-performing models capture greater share of the reward pool. Consumers and researchers gain access to state-of-the-art function-calling models validated under identical conditions, all without centralized intermediaries.

Incentive Mechanism

GroundLayer emissions follow a custom scoring split: 80 % of a miner’s score derives from evaluation accuracy (50 % single-turn, 50 % multi-turn), while 20 % stems from model longevity—assessing persistent performance across up to ten successive datasets. Rewards are minted per block at a subnet emission rate of 0.76 %, with a configurable root proportion directed to the subnet owner’s coldkey (18.3 %) and the remainder allocated among miners proportional to their scores.

Comparison to Centralized Alternatives

Unlike centralized leaderboards and proprietary API services—where model evaluation pipelines are opaque and outcomes can be altered—GroundLayer leverages Bittensor’s decentralized ledger to guarantee auditability, permissionless participation, and on-chain enforcement of rules. This architecture mitigates trust assumptions, aligns stakeholder incentives, and democratizes access to AI benchmarking beyond siloed corporate platforms.