Trace the number. Test the model.

Scope 3 is the carbon a company causes through its suppliers and customers. For most companies it is the biggest part of the footprint. It is also the hardest part to measure.

Few companies report it in full. Every report has a different layout. And a number copied out of a report rarely says where it came from.

So we came at it from both ends. Scope3Trace reads what companies report and keeps the proof. GHGBench checks how well a model can predict the numbers.

A method and a dataset. Scope3Trace pulls Scope 3 emissions out of real ESG and sustainability reports. Every number it pulls out is tied to the page that reported it. It is highly accurate on the Scope 1 to 3 totals, and on each Scope 3 category.

We used it to build a dataset. It holds Scope 3 disclosures at two levels, and each one links back to its source.

1. Step 1

   Collect the reports

   We gather sustainability and ESG reports as PDFs. OCR reads them, whatever the layout.

2. Step 2

   Find the numbers

   A language model finds the pages that report emissions. Then it rebuilds their tables.

3. Step 3

   Pull them out

   Rules and a language model work together. They pull out the Scope 1 to 3 totals and each Scope 3 category, for whole organisations and for single buildings.

4. Step 4

   Check the source

   Each number is checked against the page it came from. So you can always trace it back.

An open dataset and benchmark. Open carbon data is scattered. Each source has its own scale and its own way of scoring, so models are hard to compare.

GHGBench puts it in one place. It has two tracks, fixed train and test splits, and baselines run over several seeds. We compare results with paired bootstrap tests. That way a small gap between two models does not get oversold.

Companies.Predict a company's Scope 1+2 and Scope 3 emissions from its financial and sector data.

Buildings.Predict a building's emissions. Every source is mapped to one schema, with climate data and remote sensing embeddings added.

There are two main tests. Predict within the data the model trained on. Then predict for a region or city it has never seen. The baselines are gradient boosted trees, an MLP, an FT-Transformer, a tabular foundation model and a multimodal model.

What we found

1. 01Buildings are harder to predict than companies.
2. 02A new city hurts a model far more than swapping to a different model does.
3. 03Remote sensing helps most where the tabular models break down.
4. 04A tabular foundation model beat tuned trees on buildings across many cities. As far as we know, that is a first.