PaECTER Plant-Related Patent-Family Detector

This is the first-stage production classifier for the manuscript Identifying Plant-Related Patents: Corpus Construction and Global Patterns. It fine-tunes mpi-inno-comp/paecter to classify English patent-family title and abstract text as plant-related or not plant-related.

This release is a companion checkpoint for reproducible research on plant-related patent classification. The Apache-2.0 license metadata follows the PaECTER base-model release.

Model Details

Base model: mpi-inno-comp/paecter
Architecture: BertForSequenceClassification
Task: binary text classification
Input text: English title, blank line, English abstract
Maximum length: 512 tokens
Labels:
- no: not plant-related
- yes: plant-related
Recommended decision rule: classify as yes when p_yes >= 0.992003

This model is a supervised fine-tune of PaECTER. Please cite or link the original PaECTER model and paper when reusing this checkpoint.

Training Data

The model was trained on the repository's internal 600-family labeled DOCDB sample, split into 420 train, 90 validation, and 90 test families. The row-level labeling queue, patent text, split manifest, and prediction files are not part of this model release.

Evaluation

Internal validation and test metrics use the validation-selected threshold p_yes >= 0.992003.

Split	PR-AUC	Precision	Recall	F1
Validation	0.9989	0.9677	1.0000	0.9836
Test	0.9761	0.9091	0.9677	0.9375

An external positive-only recall check on 450 independently curated plant-related families recovered 447 families at the training-derived threshold (recall = 0.993). This external check tests recall transfer, not precision.

Intended Use

Use this model to score English title and abstract text for patent families where the goal is a high-confidence plant-related corpus. The threshold is part of the documented release behavior and should be recalibrated if used on a different corpus or label distribution.

Limitations

The labeled training set contains 600 families, so borderline behavior can shift across jurisdictions, time periods, or technical domains.
Inputs longer than 512 tokens are truncated.
The external gold check is positive-only and does not estimate external precision.
The model is trained on English title and abstract text, not full patent claims or descriptions.

Reproducibility

Reference artifacts in the companion repository:

Training report: metadata/paecter_report_2026-01-29_rerun.md
Model card source: metadata/model_card_paecter.md
Public companion repository: https://github.com/aydiet/plant-patent-classifier-reproducibility

Companion Resources

Original PaECTER base model: https://huggingface.co/mpi-inno-comp/paecter
Interactive demo Space: https://huggingface.co/spaces/aydiet/plant-patent-classifier-demo
Second-stage variety-vs-technology classifier: https://huggingface.co/aydiet/plant-patent-paecter-variety-technology
Public companion repository: https://github.com/aydiet/plant-patent-classifier-reproducibility
Manuscript context: Identifying Plant-Related Patents: Corpus Construction and Global Patterns

Example

from transformers import AutoModelForSequenceClassification, AutoTokenizer
import torch

model_id = "aydiet/plant-patent-paecter-plant-detector"
text = "Drought tolerant maize plant\n\nA maize plant with improved drought tolerance is provided."

tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForSequenceClassification.from_pretrained(model_id)
inputs = tokenizer(text, return_tensors="pt", truncation=True, padding="max_length", max_length=512)

with torch.no_grad():
    probs = torch.softmax(model(**inputs).logits, dim=-1)[0]

p_yes = float(probs[model.config.label2id["yes"]])
label = "yes" if p_yes >= 0.992003 else "no"
print(label, p_yes)