specter2-review-rationale

Model: specter2-review-rationale

The model snsf-data/specter2-review-rationale is based on the allenai/specter2_base model and fine-tuned for a binary classification task. In particular, the model is fine-tuned to classify if a sentence from SNSF grant peer review report is addressing the following aspect:

Does the sentence provide rationale supporting the positive or negative statement?

The model was fine-tuned based on a training set of 2'500 sentences from the SNSF grant peer review reports, which were manually annotated by multiple human annotators via majority rule. The fine-tuning was performed locally without access to the internet to prevent any potential data leakage or network interference. The following setup was used for the fine-tuning:

• Loss function: cross-entropy loss
• Optimizer: AdamW
• Weight decay: 0.01
• Learning rate: 2e-5
• Epochs: 3
• Batch size: 10
• GPU: NVIDIA RTX A2000

The model was then evaluated based on a validation set of 500 sentences, which were also manually annotated by multiple human annotators via majority rule. The resulting macro-average F1 score: 0.71 was achieved on the validation set. The share of the outcome label amounts to 18.4%.

The fine-tuning codes are open-sourced on GitHub: https://github.com/snsf-data/ml-peer-review-analysis .

Due to data privacy laws no data used for the fine-tuning can be publicly shared. For a detailed description of data protection please refer to the data management plan underlying this work: https://doi.org/10.46446/DMP-peer-review-assessment-ML. The annotation codebook is available online: https://doi.org/10.46446/Codebook-peer-review-assessment-ML.

For more details, see the following paper published in Quantitative Science Studies:

A Supervised Machine Learning Approach for Assessing Grant Peer Review Reports

by Gabriel Okasa, Alberto de León, Michaela Strinzel, Anne Jorstad, Katrin Milzow, Matthias Egger, and Stefan Müller, available open-access: https://doi.org/10.1162/QSS.a.23 .

How to Get Started with the Model

The model can be used to classify sentences from grant peer review reports for addressing rationale when making positive or negative statements.

Use the code below to get started with the model.

# import transformers library
import transformers

# load tokenizer from specter2_base - the base model
tokenizer = transformers.AutoTokenizer.from_pretrained("allenai/specter2_base")

# load the SNSF fine-tuned model for classification of rationale in review texts
model = transformers.AutoModelForSequenceClassification.from_pretrained("snsf-data/specter2-review-rationale")

# setup the classification pipeline
classification_pipeline = transformers.TextClassificationPipeline(
    model=model,
    tokenizer=tokenizer,
    return_all_scores=True
)

# prediction for an example review sentence addressing rationale
classification_pipeline("Given the comments above, this reviewer considers that the proposed methods are well suited to provide key answers to the proposed hypotheses, particularly because these methods and models have provided the background knowledge justifying the reverse engineering approach.")

# prediction for an example review sentence not addressing rationale
classification_pipeline("There are currently several activities on an international level that have identified the issue and activities are underway.")

Model Limitations

• Human Assessment Required: This model should not be used for automatic classification of grant peer review reports without human oversight.
• Limited Training Data: The model was fine-tuned on a limited sample of 2,500 annotated sentences. Therefore, its classification accuracy should be critically evaluated before deployment.
• Specific Training Data: The training data consists of a random sample of SNSF grant peer review reports. As such, the model's external validity to other datasets may be limited.

Disclaimer

The classification accuracy of this model is notably low. Therefore, any deployment should be approached with caution and thorough consideration.

Since Rationale was only considered for Positive or Negative statements in the annotation, the predictions of the snsf-data/specter2-review-rationale model should be adapted in conjunction with the prediction of the models snsf-data/specter2-review-positive and snsf-data/specter2-review-negative as follows:

• if a sentence's prediction is either Positive or Negative, retain its Rationale prediction
• if a sentence's prediction is neither Positive nor Negative, disregard its Rationale prediction.

Citation

BibTeX:

@article{okasa2025supervised,
  title={A supervised machine learning approach for assessing grant peer review reports},
  author={Okasa, Gabriel and de Le{\'o}n, Alberto and Strinzel, Michaela and Jorstad, Anne and Milzow, Katrin and Egger, Matthias and M{\"u}ller, Stefan},
  journal={Quantitative Science Studies},
  volume={6},
  pages={1189--1214},
  year={2025},
  publisher={MIT Press 255 Main Street, 9th Floor, Cambridge, Massachusetts 02142, USA}
}

APA:

Okasa, G., de León, A., Strinzel, M., Jorstad, A., Milzow, K., Egger, M., & Müller, S. (2025). A supervised machine learning approach for assessing grant peer review reports. Quantitative Science Studies, 6, 1189-1214.

Model Card Authors

Gabriel Okasa, Alberto de León, Michaela Strinzel, Anne Jorstad, Katrin Milzow, Matthias Egger, and Stefan Müller

Model Card Contact

[email protected]