Instructions to use CobuildX-AI/cbx-bfs-esg-topic-classifier-17M with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- Transformers
How to use CobuildX-AI/cbx-bfs-esg-topic-classifier-17M with Transformers:
# Use a pipeline as a high-level helper from transformers import pipeline pipe = pipeline("text-classification", model="CobuildX-AI/cbx-bfs-esg-topic-classifier-17M")# pip install -U transformers accelerate # Load model directly from transformers import AutoTokenizer, AutoModel tokenizer = AutoTokenizer.from_pretrained("CobuildX-AI/cbx-bfs-esg-topic-classifier-17M") model = AutoModel.from_pretrained("CobuildX-AI/cbx-bfs-esg-topic-classifier-17M", device_map="auto") - Notebooks
- Google Colab
- Kaggle
cbx-bfs-esg-topic-classifier-17M
An English ESG topic classifier. Give it a passage (a paragraph from an annual report, a proxy statement, a regulator's press release, a news item or a credit memo) and it returns:
- topics: any of 15 environmental, social and governance topics the passage substantively discusses, or none;
- primary: the main topic (
N00when there is none); - orientation: whether the passage reports
adverseevents (harm, violations, allegations, penalties),positiveconcrete measures or results,neutraldescription (policies, risk statements, regulation explained), ormixed.
It is a 17M-parameter Ettin encoder (jhu-clsp/ettin-encoder-17m) with three heads. A 68M version is also available.
Licence: Apache-2.0. Read Intended use and disclaimer before using it.
Topics
A topic counts only when the passage says something substantive about it, not when it is merely named. Test F1 is on held-out companies and 2025-2026 agency releases.
| Code | Topic | Covers | Test F1 |
|---|---|---|---|
E01 |
Emissions and climate | GHG emissions, targets, decarbonisation, carbon pricing, climate-change risk | 0.772 |
E02 |
Energy use and transition | energy efficiency, renewable capacity or purchase, energy-mix shifts | 0.769 |
E03 |
Water | water use, withdrawal, scarcity, conservation, wastewater capacity | 0.859 |
E04 |
Pollution and waste | toxic releases, air permits, spills, contamination, waste, recycling, remediation | 0.861 |
E05 |
Nature and land use | habitats, biodiversity, protected species, deforestation, land restoration | 0.766 |
S01 |
Workers' rights and conditions | wages, hours, discrimination, harassment, unions, forced or child labour, DEI with substance | 0.731 |
S02 |
Workplace health and safety | worker injuries and fatalities, exposure, injury rates, safety programmes | 0.824 |
S03 |
Customer protection and fair treatment | mis-selling, deceptive fees, unfair lending or collection, consumer fraud by a company | 0.740 |
S04 |
Product safety and quality | defective or unsafe products, recalls, food / drug / device safety | 0.857 |
S05 |
Privacy and cybersecurity | personal-data use, breaches, ransomware, cybersecurity programmes | 0.869 |
S06 |
Community impact and access | effects on communities, displacement, access to essential services and finance | 0.800 |
G01 |
Board and executive accountability | board independence and oversight, executive pay design, related parties, shareholder rights | 0.794 |
G02 |
Ethics and anti-corruption | bribery, fraud, insider trading, money laundering, sanctions evasion, antitrust | 0.771 |
G03 |
Reporting integrity and transparency | misstatements, restatements, material weaknesses, auditor issues, greenwashing | 0.739 |
G04 |
Responsible finance and investment | ESG screening in lending or investing, green / social bonds, financed emissions, stewardship | 0.692 |
Quick start
pip install -r requirements.txt
python example.py
from cbx_esg import ESGClassifier
clf = ESGClassifier(".") # folder with the model files; runs on CPU or GPU
r = clf.classify("Fenwick Foods agreed to pay a civil penalty to resolve allegations that it failed to report "
"injuries at its packaging plant ...")
r["topics"], r["primary"], r["orientation"] # (['S02'], 'S02', 'adverse')
r["topic_scores"] # score per topic; topics are those >= the threshold
classify also takes a list of passages. The default threshold (0.63) was tuned on validation for
micro F1; pass threshold= to ESGClassifier to trade recall for precision.
Input
- English prose, one passage at a time. Training passages were about 75-250 words.
- At most 512 tokens are read; longer text is cut (
"truncated": True). Split long documents into paragraphs. - Tables, figures, headers and lists of numbers are outside what it was trained on.
Evaluation
Topic micro F1, macro F1 over the 15 topics, the share of no-topic passages given a topic (false alarm), primary accuracy and orientation accuracy. TF-IDF + logistic regression trained on the same data is shown for reference. No evaluation passage, company or document was used in training.
| Test set | Micro F1 | Macro F1 | False alarm | Primary acc. | Orientation acc. | TF-IDF micro F1 |
|---|---|---|---|---|---|---|
| Held-out companies and 2025-2026 agency releases (1,761 passages) | 0.799 | 0.790 | 0.103 | 0.847 | 0.916 | 0.810 |
| Two regulators never seen in training: UK competition authority, US product-safety agency (363) | 0.695 | 0.557 | 0.449 | 0.614 | 0.573 | 0.619 |
| US Federal Register rule text, a genre absent from training (439) | 0.664 | 0.633 | 0.174 | 0.676 | 0.950 | 0.705 |
Agency and court names were masked in half of the training passages; masking them in the test sets changes test micro F1 by 0.001.
Training data
- 12,030 passages for training and validation: 10,870 real English passages, 2019-2026 (SEC EDGAR 10-K, 20-F and proxy filings of about 1,200 companies; releases of the U.S. Department of Justice, SEC, Federal Reserve Board and FDIC, GOV.UK and the European Commission), plus 1,160 generated passages with invented names for thin topics and hard negatives.
- Labels follow a written annotation guide (topic definitions, boundary rules, primary and orientation rules). Each passage was labelled independently by two LLMs (gpt-5.4-mini and Claude Haiku 5.5); disagreements, plus a 10% audit of agreements, were decided by Claude Sonnet 5.5. A further set of real passages was chosen where an earlier model was least certain. Labels were not reviewed by people row by row and contain errors.
Limitations
- Rule-making text. On U.S. Federal Register rule text, a genre not in the training data, micro F1 is 0.66 and 17% of no-topic passages get a topic. Board / governance (G01) and reporting integrity (G03) are weakest there.
- Unseen publishers. On two regulators never seen in training, macro F1 is 0.56; topics those regulators rarely discuss are unreliable.
- Responsible finance (G04) is the weakest topic on the main test set (0.69).
- English, 2019-2026, mostly U.S. / UK / EU sources. Other jurisdictions, languages, sectors and writing styles (social media, ratings reports, scanned documents) are untested.
- Topic, not truth. The model says what a passage is about. It does not check whether claims are true, and it does not measure ESG performance.
Intended use and disclaimer
Intended use: research; tagging and searching ESG-related passages in reports, filings, news and internal documents; routing passages for review by a person; aggregate statistics over many documents; testing and comparing ESG text tools.
Not intended for, and not supported:
- Investment, lending, credit or insurance decisions. Do not use the model or its outputs as the basis, or a decisive input, for decisions to invest, lend, price, underwrite, rate or exclude any company, issuer, borrower or security. Its outputs are not an ESG rating, score or assessment of any entity.
- Legal, regulatory or compliance determinations, including whether any disclosure, taxonomy, reporting or anti-greenwashing requirement is met. Its outputs are not legal, regulatory, financial, investment or compliance advice and may be wrong.
- Statements about real companies or people. A topic or an
adverseorientation means only that a passage discusses that subject; it is not a finding that anyone did anything. Do not publish or act on model outputs as claims about a named company or person without checking the source. - Decisions about individuals, such as employment, credit or access to services.
No responsibility for use. The model is provided "as is", without warranties or conditions of any kind, and the authors and Cobuild Tech Private Limited (CobuildX) are not liable for any use of it (Apache License 2.0, sections 7 and 8). You alone are responsible for how you use the model and its outputs, for any decision taken with them, and for complying with the law that applies to you. The model runs on your systems; CobuildX does not receive your data. Scores on this page are measured on specific test sets and are not a guarantee of performance on your text.
Files
| File | |
|---|---|
model.safetensors, config.json, tokenizer.json, tokenizer_config.json |
the model: Ettin encoder with topic, primary and orientation heads |
esg_config.json |
topic codes and names, labels, threshold, input limit, base model, scores |
cbx_esg.py, example.py, requirements.txt |
loader and example |
LICENSE, NOTICE |
Apache-2.0 licence, third-party notices and data attributions |
checksums.sha256 |
SHA-256 of every file |
Contact
Questions: hello@cobuildx.ai. Release v0.1 (2026-11-18).
- Downloads last month
- 17
Model tree for CobuildX-AI/cbx-bfs-esg-topic-classifier-17M
Base model
jhu-clsp/ettin-encoder-17m