spice
Temporal ML research pipeline for EVM fee-timing decisions.
Temporal fee-decision deep learning pipeline.
// overview
SPICE turns blockchain block history into supervised temporal decision problems. It acquires canonical EVM block data, builds fee-dynamics features, trains neural sequence models, decodes predictions into timing offsets, and evaluates those decisions with replay-style economic metrics.
The project is structured as a research and operations system rather than a notebook. Typed YAML surfaces define experiments, workflows persist corpora, Optuna studies, model artifacts, evaluations, and benchmark result indexes.
// how it works
A Typer CLI and benchmark runner resolve configuration surfaces into acquisition, training, tuning, prediction, and evaluation workflows.
The pipeline builds or loads a corpus, constructs feature tables, compiles temporal examples, tensorizes sequences, trains model families, decodes outputs into fee-delay decisions, then scores them through evaluator contracts.
// capabilities
EVM corpus acquisition
Collects block-history data for Ethereum, Polygon, and Avalanche through RPC providers.
Sequence model comparison
Runs LSTM, Transformer, and Transformer-LSTM families through repeatable training and tuning loops.
Economic replay evaluation
Scores decoded timing offsets with temporal replay evaluators instead of only model-loss metrics.
Durable experiment state
Stores corpora, studies, artifacts, evaluations, catalogs, and transfer outputs under explicit roots.