vector-institute/dia-state-lab-2026
DIA lab footprint table This dataset is the rollup table for the Data & Impact Accounting (DIA) lab demo. It indexes the training footprint (energy, carbon, water) and lineage of a set of demo models trained across A100 / A40 / CPU hardware. It is produced by ingesting each model's dia_report card and is read by the DIA Gradio dashboard. It stores metadata only — no model weights. Files nodes.parquet — one flat row per model (browsable in the Dataset Viewer):… See the full description on the dataset page: https://huggingface.co/datasets/vector-institute/dia-state-lab-2026.
DIA lab footprint table
This dataset is the rollup table for the Data & Impact Accounting (DIA) lab demo. It indexes the training footprint (energy, carbon, water) and lineage of a set of demo models trained across A100 / A40 / CPU hardware.
It is produced by ingesting each model's dia_report card and is read by the DIA Gradio dashboard. It stores metadata only — no model weights.
Files
- `nodes.parquet` — one flat row per model (browsable in the Dataset Viewer): energy/carbon/water intervals, data-quality tier, GPU, GPU-hours, region, lineage.
- `state.json` — the nested source of truth the dashboard loads.
How the rollup works
Given a base model, the dashboard builds the lineage as a directed graph and takes the base plus all its descendants as the family, then:
- Sums incremental footprints — each model logs only its own training delta; the family total is the subtree sum.
- Dedupes the DAG — a merged/shared model is counted once.
- Reports coverage, not a bare total — totals are a lower bound at low disclosure.
- Keeps provenance separate —
measuredvsestimatedvsimputed.
Related
- Toolkit / paper: VectorInstitute/ai-impact-accounting
- Models: the
DIA-MVP/*-a100,*-a40,*-cpurepos ingested here.
