ASTRA EPC trained model bundles
This private repository stores 129 fitted model bundles and their evaluation metrics. It includes the 18-target CatBoost, TabICL v2 and TabPFN v3 EPC benchmark (54 bundles), XGBoost and LightGBM EPC comparisons (36), seven missing-fact estimators across five model families (35), and four EPC-linked property-value models. See epc/runs/<task>/<family>/metrics.json for per-run metrics and checksums. The audited expansion report is epc/expansion-evaluation.json.
Preparation, feature contracts, validation evidence and held-out results are documented in the project repository. The pipeline code commit used for the expansion is 3ece34b.
Scope and limits
These are research and evaluation artifacts, not production-certified valuations or engineering assessments. EPC recommendation models predict assessor-listed categories; their scores do not establish physical feasibility, savings, or causal retrofit value. Missing-fact models estimate categories from other EPC/property fields and should surface uncertainty, especially for rare classes.
Property valuation uses an observational UCL/GLA linked Land Registry and EPC dataset, with non-commercial research terms stated by the publisher. It covers England and Wales through October 2024, not current 2026 prices. The source has no direct EPC lodgement-date field; date filtering and property grouping use inferred LMK_KEY tokens and are not verified against source dates or a UPRN. This model is not a consumer valuation and cannot estimate causal retrofit uplift.
The repo contains fitted bundles, model metadata and aggregate metrics, but no raw EPC/property records, addresses, or test-prediction CSVs. TabPFN v3 carries a non-commercial model-weights licence. This repository is private; do not redistribute TabPFN bundles or use them commercially without checking and obtaining required rights. The linked valuation data are for non-commercial research only unless licensing is separately established.
Loading
Each serialized model.joblib contains a fitted estimator, feature transformation and schema. Load with joblib.load, then use the matching project inference code and pinned environment from configs/ml/runtime-versions.json. Bundle checksums and reload verification are recorded in each run report.