A demo model from the Data & Impact Accounting (DIA) lab. It performs instruction-tuning (LoRA adapter) via LoRA (PEFT), with the base model meta-llama/Llama-3.2-3B-Instruct, trained on NVIDIA A100.

The point of this repo is not the model itself but its dia_report — a standardized record of the energy, carbon, and water used to train it, embedded in this card's metadata.

This footprint feeds the DIA dashboard, which rolls up a base model and all its derivatives to show the cumulative carbon, water, and energy cost of a model family.

Training footprint

Metric	Value
Hardware	1× NVIDIA A100-SXM4-80GB
Compute	0.3014 GPU-hours
Energy	0.1082 (measured) kWh
Carbon	0.007 (measured) kgCO₂eq
Water	0.195–0.433 (estimated-from-default-wue) L
Grid region	ca-on

Energy and carbon are measured with CodeCarbon; water is estimated from a default water-usage-effectiveness range. Carbon uses the local grid's intensity (Ontario, ~0.03 kgCO₂eq/kWh).

Reproduce

REPO=DIA-MVP/llama32-3b-lora-a100 python scripts/train_llama_lora.py

DIA-MVP/llama32-3b-lora-a100

Llama 3.2 3B Instruct — LoRA (NVIDIA A100)

Training footprint

Reproduce

Links

DIA-MVP/
llama32-3b-lora-a100