Samarth-27/carbon-mrv-invoice-emissions

Synthetic Carbon MRV Invoice-to-Emissions Dataset

A synthetic dataset that models the core pipeline used by carbon Measurement, Reporting & Verification (MRV) platforms: turning a business document line item (invoice, fuel receipt, electricity bill, freight charge) into a GHG Protocol Scope 1 / 2 / 3 classification and a calculated emissions value.

It was built as reference dataset for learning and prototyping — specifically for training/evaluating models that do:

• Scope classification: given a line-item description (and/or HSN code), predict whether it's Scope 1, 2, or 3, and which GHG Protocol category it falls under.
• Emissions estimation: given category, quantity, unit and country, predict calculated_emissions_kgco2e.
• OCR/extraction robustness: line_item_raw_text includes a noisy (typo-injected) variant of ~35% of rows, to simulate imperfect OCR output vs. line_item_clean_text.

⚠️ Important: synthetic data, not compliance-grade

This dataset is fully synthetic. Invoices, vendor names, quantities and prices are randomly generated — they are not real transactions from any company. Emission factors are approximate reference values pulled from public methodologies (see Sources below), not the latest official factors for any specific jurisdiction or reporting year. Do not use this for actual regulatory disclosure, CBAM filings, or carbon credit issuance. It's meant for model training, coursework, portfolio projects, and prototyping carbon-tech pipelines.

Dataset structure

1,500 rows across ~600 synthetic invoices, split 80/10/10 into train / validation / test (also provided as separate CSV files).

Column	Type	Description
id	int	Row identifier
invoice_id	int	Groups 1-4 line items into a synthetic invoice
vendor_name	string	Synthetic vendor name
date	string (YYYY-MM-DD)	Synthetic invoice date (2025-2026)
country	string	One of 10 Asian markets (India, Bangladesh, Vietnam, Indonesia, Philippines, Thailand, Malaysia, Pakistan, Singapore, Sri Lanka)
industry	string	Steel, Textile & Apparel, Chemicals, Logistics & Freight, General Manufacturing, Food Processing, Cement & Building Materials
line_item_raw_text	string	Invoice-style description, ~35% with OCR-like noise (typos, digit/letter substitution)
line_item_clean_text	string	Canonical clean description (extraction "ground truth")
hsn_code	string	Illustrative HSN/SAC-style code for the item category
category	string	17 categories spanning fuels, electricity, purchased goods, freight, travel, waste
ghg_scope	int (1/2/3)	GHG Protocol scope — the main classification target
ghg_protocol_category	string	Finer-grained GHG Protocol category (e.g. "Scope 3 - Upstream Transportation & Distribution")
quantity	float	Quantity in unit
unit	string	litres, kg, m3, kWh, tonne-km, passenger-km
unit_price	float	Synthetic unit price
currency	string	Local currency code for the country
total_amount	float	quantity * unit_price
emission_factor	float	Emission factor applied
emission_factor_unit	string	e.g. kgCO2e/litre, kgCO2e/kWh
emission_factor_source	string	Reference methodology/source for the factor used
calculated_emissions_kgco2e	float	quantity * emission_factor — the regression target
split	string	train / validation / test

Category coverage

Scope 1 (direct combustion): Diesel, Petrol, LPG, Furnace Oil, Industrial Coal, Natural Gas Scope 2 (purchased electricity): grid electricity, factor varies by country (IEA 2023-style country grid intensities) Scope 3 (value chain): Steel, Cement, Cotton Yarn, Plastic Packaging (purchased goods) · Road/Rail/Sea/Air Freight (upstream transport) · Business Air Travel · Waste Disposal

Roughly 34% Scope 1, 6% Scope 2, 60% Scope 3 — deliberately weighted toward Scope 3 since that's typically 65-90% of a company's real footprint and the hardest to classify well.

Sources for emission factors (approximate, illustrative)

• IPCC 2006 Guidelines for National GHG Inventories, Vol. 2 (fuel combustion factors)
• DEFRA 2024 UK Government GHG Conversion Factors (materials, waste, travel)
• IEA 2023 country grid emission intensities (electricity)
• GLEC Framework v3 (freight/logistics emission factors)
• worldsteel Association and Global Cement & Concrete Association (cradle-to-gate material factors, approximate)
• Higg Materials Sustainability Index / Textile Exchange (textile material factors, approximate)

Example use cases

from datasets import load_dataset

ds = load_dataset("Samarth-27/carbon-mrv-invoice-emissions")

# Task 1: predict GHG scope from the line-item text
X = ds["train"]["line_item_clean_text"]
y = ds["train"]["ghg_scope"]

# Task 2: predict emissions from structured fields
features = ds["train"].select_columns(
    ["category", "country", "quantity", "unit", "emission_factor"]
)
target = ds["train"]["calculated_emissions_kgco2e"]

# Task 3: OCR robustness - compare raw vs clean text
noisy = [r for r in ds["train"] if r["line_item_raw_text"] != r["line_item_clean_text"]]

Reasonable baseline models to try: TF-IDF + logistic regression or a small fine-tuned transformer for scope classification; gradient boosting (XGBoost/LightGBM) for the emissions regression task, since it's mostly quantity x emission_factor with categorical structure.

Note on difficulty: classifying scope from line_item_clean_text is close to trivial (category names are embedded in the description, so a TF-IDF + logistic regression baseline hits ~100% accuracy). The more realistic and interesting task is classifying from line_item_raw_text (the noisy/OCR-like variant) or from partial/ambiguous descriptions — that's where a model actually has to generalize, closer to what a real MRV extraction pipeline has to handle.

How this was built

Generated programmatically (generate_dataset.py, included in this repo) using a fixed random seed for reproducibility. No real company or personal data was used at any stage.

License

Released under MIT license. Free to use, modify, and redistribute for research, learning, and prototyping.

Citation

If you use this dataset in a project, a simple reference back to this repo is appreciated. This is an independently built reference dataset and is not affiliated with or endorsed by Senseible, Biocog, or any carbon MRV vendor.