InnoBERT model card

Python package and usage guide: https://github.com/mustafahci/InnoBERT

Model description

InnoBERT is a BERT-base sequence classifier fine-tuned from yiyanghkust/finbert-pretrain. It has eight sigmoid outputs in this fixed order:

1. inno_product
2. inno_process
3. inno_organizational
4. inno_marketing
5. inno_businessmodel
6. inno_sustainability
7. inno_AI
8. inno_uncategorized

The supplied checkpoint is a BertForSequenceClassification model with 12 layers, hidden size 768, 12 attention heads, 30,873 vocabulary items, and an 8-by-768 classifier head.

InnoBERT classifies the nature of submitted text; it does not determine whether the language is novel. In the measurement procedure developed by Ahci and Joos (2026), novelty is established first by comparing terms with prior economy-wide business disclosures. InnoBERT is then used to classify the nature of those novel terms. When used independently, its output should be interpreted as a semantic category assignment rather than, by itself, evidence of innovation.

The Python package reports reader-facing granular labels without the internal inno_ prefix and maps process, organizational, marketing, and business-model predictions to the main category business_process. This reporting hierarchy does not alter the model outputs or thresholds.

The innovation-type framework is informed by the Oslo Manual 2018. Sustainability, AI, and the uncategorized outcome are additional reporting categories used by InnoBERT rather than being presented as separate official Oslo Manual innovation types.

Intended use

The primary use is multi-label semantic classification of candidate terms extracted from business disclosures. The validated training input includes an industry, year, and marked term. The package also supports sentence and paragraph processing as documented extensions. The model does not reconstruct the economy-wide novelty stage or automatically create firm-year innovation measures.

It is not intended for individual-level decisions, high-stakes automated decisions, or claims about realized innovation without separate validation.

In noun-chunk mode, surrounding statements are removed before classification. Negation, prospective language, attribution to competitors or other actors, and evidence of actual adoption are therefore outside the noun-chunk classifier's output. A label identifies the semantic category associated with an extracted phrase; it does not establish what the focal firm did.

Preprocessing and thresholds

The noun-chunk mode preserves internal hyphens, removes specified generic modifiers, lemmatizes noun heads, splits selected coordinated noun phrases, limits phrases to four terms, removes subphrases, and optionally removes filer-name tokens. Its corpus-derived stoplist is frozen in the Python package. Identical processed phrases are deduplicated within each source document but retained separately across source documents. Source and unit identifiers preserve document-level lineage; they are not mention counts or character offsets.

Default thresholds are 0.65, 0.45, 0.55, 0.55, 0.45, 0.50, 0.50, and 0.25 in the label order above.

For contextual term and noun-chunk inputs, the default uncategorized rule is a conservative rejection gate: an uncategorized score at or above 0.25 suppresses substantive category assignments, even when another category has a higher score. This prioritizes detection of potentially irrelevant terms and reduces false-positive category assignments at the cost of lower recall. If no substantive category reaches its threshold, the output falls back to uncategorized rather than returning an empty label list. Sentence and paragraph modes use uncategorized only as this fallback. Industry and year are part of the contextual model input and must be defined consistently across a research sample.

Limitations

• The training data are not distributed with this release.
• Sentence and paragraph inputs differ from the contextual term format used for training.
• Windowed paragraph scores use category-wise maxima and are not calibrated paragraph-level probabilities.
• Noun-chunk predictions do not capture negation, adoption status, timing, or attribution from the surrounding sentence.
• Results can vary if noun chunks are extracted with a different spaCy model or version.
• Disclosure language is not equivalent to realized innovation capability or outcomes.
• The training data and annotations are not distributed, so users cannot reconstruct training from this release.
• Applying the classifier outside the disclosure setting, industries, or years represented in development requires separate validation.
• Research accuracy for a new corpus or construct should be evaluated against a labelled sample from that intended setting.

Citation

Ahci, Mustafa and Joos, Philip, Beyond Invention: The Composition and Economic Relevance of Innovation-Related Capabilities (Updated September 1, 2026). Available at SSRN: https://ssrn.com/abstract=4797745 or http://dx.doi.org/10.2139/ssrn.4797745.

OECD/Eurostat (2018), Oslo Manual 2018: Guidelines for Collecting, Reporting and Using Data on Innovation, 4th Edition, The Measurement of Scientific, Technological and Innovation Activities, OECD Publishing, Paris/Eurostat, Luxembourg. https://doi.org/10.1787/9789264304604-en.

Upstream attribution

InnoBERT was independently fine-tuned from Huang, Wang, and Yang's FinBERT checkpoint and is not affiliated with or endorsed by the original FinBERT or BERT developers.

License

The InnoBERT fine-tuned weights and accompanying code are provided under Apache License 2.0. The official FinBERT GitHub project is Apache-2.0 licensed and directly identifies the pretrained FinBERT models; however, yiyanghkust/finbert-pretrain does not independently declare license metadata on its Hugging Face model card. This provenance qualification is disclosed rather than obscured. Users requiring formal legal certainty—especially for commercial redistribution—should independently confirm that the upstream license covers the checkpoint weights.