Implementation and Evaluation of a Blockchain and Decentralized Identity-Based Federated Learning Prototype for GPT-2 Fine-Tuning

Large language model development depends on large-scale training data, yet centralized collection raises privacy risks and concentrates control among a few organizations. Federated learning addresses part of this problem but retains a central aggregator, cannot readily verify contributions, and lacks sustained incentives. Blockchain integration resolves coordination and transparency, yet participant identity remains exposed to sybil attacks. This study designs, implements, and evaluates a prototype platform combining these three technologies. The prototype uses GPT-2 Small with 10,000 Alpaca samples partitioned Non-IID across five clients, the FederatedHub contract on Ethereum, and IPFS as off-chain storage. Evaluation covers 5 rounds across 3 seeds, 9 attack variants, a sybil attack using 10 identities, and two-attacker collusion. The delta-based scoring function yields scores from 27.3 to 73.0 correlating 0.920 to 0.991 with data volume, perplexity falls from 24.6722 to 7.7798, and ROUGE-L rises by 106.6%. The gate catches 8 of 9 variants, yet one variant that positions itself inside the acceptance range passes every trial while capturing 95.2% of legitimate client rewards. All sybil identities are admitted, and gas cost limits the system to 15 or 16 participants. The weakness lies in the identity and incentive layers, not in model quality.