TrustCompute: Result Validation, Reputation-Aware Scheduling and Quality-Based Payments for Outsourced LoRA Fine-Tuning

Abstract

Outsourced model fine-tuning requires reliable task allocation, result validation, and incentives for quality. We present TrustCompute, a compute-sharing platform integrating two-stage adapter verification, reputation-aware scheduling, quality-based payments, and auditable ledger records for LoRA fine-tuning. Our evaluation combines synthetic workloads, real GPU-based fine-tuning, and a local Ethereum Virtual Machine (EVM) environment. Across 440 validation cases, all 400 invalid submissions were rejected and all 40 valid submissions were accepted. With 50% unreliable workers, reputation-aware scheduling achieved 100% task completion and reduced the mean failed-attempt time by 89% relative to round-robin scheduling. An integration experiment completed 120 real fine-tuning tasks. Additional simulations showed that identity resets can exploit exploration and that quality proportional payments can improve strategic workers’ delivered quality. These findings support jointly designing verification, scheduling, and payment mechanisms. TrustCompute verifies compliance with result acceptance criteria, but does not prove adherence to a prescribed training procedure.