New Dataset of Performance Metrics, Device Telemetry, and Blockchain Audit Logs from a Federated Learning Testbed on Raspberry Pi Devices for HAR

This dataset contains new data produced by federated learning training experiments. It is not the original MHEALTH or UCI-HAR datasets. Those public benchmarks were used only as the training source. The deposit holds the raw output and the aggregated metrics from 40 federated learning experiments that ran on a physical testbed of 10 Raspberry Pi 4 devices. The aim was to test whether on-chain auditing can be added to real edge-device federated learning without lowering model accuracy or adding too much overhead. In every round the system records the round on Base mainnet, an Ethereum Layer 2 network, and stores the model artefacts on IPFS. Two operational variants are compared on two activity recognition benchmarks, MHEALTH and UCI-HAR. The baseline variant runs the blockchain steps synchronously. The optimised variant uploads to IPFS in the background and caches gas parameters. Model accuracy stays the same across both variants, within half a percentage point. The optimised variant roughly halves the median blockchain overhead per round on both benchmarks. The main source of variation is the time the public network takes to confirm a transaction. Chain integrity checks passed for all 40 sessions. The dataset covers 1,240 transactions in total, 620 per benchmark, and every transaction is listed by its hash with a direct Basescan link in the pin map files. The same contract address also carries transactions from other runs, so its total on Basescan is 4963 transactions. The deposit holds 12 flat-table CSV files with 15,280 data rows in total. They cover round-level metrics, per-device hardware telemetry, per-client evaluation records, session summaries, millisecond-level blockchain and IPFS timings for both benchmarks, and a blockchain to IPFS pin audit map. Two zip archives hold the complete raw output of all 40 sessions. Each archive holds the 20 raw session folders for one benchmark, and every folder contains its results log in NDJSON form, its configuration file, its blockchain and server performance logs, and eleven confusion matrix images. A field-level data dictionary named DATA_DICTIONARY.md gives the row count and columns for every file. The smart contract is live on Base mainnet at https://basescan.org/address/0xBE3FeAC76293711C5D32426860303AfE24cdf527 Code and aggregation scripts are at https://github.com/bmukhambetiyar/FLBL2. Version 2: completes the UCI-HAR experiments. All 20 UCI-HAR sessions are now complete, each with full blockchain provenance and a performance log. This version adds perf_timings_ucihar.csv and raises the CSV count from eleven to twelve. The per-client records are now complete for UCI-HAR, with all ten clients present in every round. The UCI-HAR raw archive and the data dictionary were rebuilt to match the new data.