SetFit-FinBERT1-Avg-challenge

Detect Actions in Asynchronous Conversation Comments

SetFit with TurkuNLP/bert-base-finnish-cased-v1

This is a SetFit model that can be used for Text Classification of actions in asynchronous conversation. This particular model detects if a comment includes a challenge or not. The configuration of the model is that the model is based on averaged annotations (from 3 annotators). Metric evaluations are based on conservative ground truth (see paper). This SetFit model uses TurkuNLP/bert-base-finnish-cased-v1 as the Sentence Transformer embedding model (using word embeddings). A LogisticRegression instance is used for classification.

The model has been trained using an efficient few-shot learning technique that involves:

1. Fine-tuning a Sentence Transformer with contrastive learning.
2. Training a classification head with features from the fine-tuned Sentence Transformer.

Model Details

Model Description

• Model Type: SetFit
• Sentence Transformer body: TurkuNLP/bert-base-finnish-cased-v1
• Classification head: a LogisticRegression instance
• Maximum Sequence Length: 512 tokens
• Number of Classes: 2 classes (action present yes/no)

<!-- - **Training Dataset:** [Unknown](https://huggingface.co/datasets/unknown) --> <!-- - **Language:** Unknown --> <!-- - **License:** Unknown -->

Model Sources

• Repository: GitHub
• Paper: Paakki, H., Toivanen, P. and Kajava K. (2025). Implicit and Indirect: Detecting Face-threatening and Paired Actions in Asynchronous Online Conversations. Northern European Journal of Language Technology (NEJLT), 11(1), pp. 58-83.

Model Labels

Evaluation

Metrics

<!--| **accuracy** | 0.7973 |-->

| 5-fold cross-validated F1 | 0.54 |

Uses

Direct Use for Inference

First install the SetFit library:

pip install setfit

Then you can load this model and run inference.

from setfit import SetFitModel

# Download from the 🤗 Hub
model = SetFitModel.from_pretrained("Finnish-actions/SetFit-FinBERT1-Avg-challenge")
# Run inference
preds = model("Kohta on lisää lapsia sairaalassa koronan vuoksi ☹")

Downstream Use

NB. This model has been trained on data coming from Finnish language asynchronous conversations under crisis related news on Facebook. This specific model has been trained to detect whether a comment includes a question or not. It reflects only one of our annotators' label interpretations, so the best use of our models (see our paper) would be to combine a set of models we provide on our Huggingface (Finnish-actions), and use a model ensemble to provide label predictions. It needs to be noted also that the model may not be well applicable outside of its empirical context, so in downstream applications, one should always conduct an evaluation of the model applicability using manually annotated data from that specific context (see our paper for annotation instructions).

Out-of-Scope Use

Please use this model only for action detection and analysis. Uses of this model and the involved data for generative purposes (e.g. NLG) is prohibited.

Bias, Risks and Limitations

Note that the model may produce errors. Due to the size of the training dataset, model may not generalize very well even for other novel topics within the same context. Note that model predictions should not be regarded as final judgments e.g. for online moderation purposes, but each case should also be regarded individually if using model predictions to support moderation. Also, the annotations only reflect three (though experienced) annotators' interpretations, so there might be perspectives on data intepretation that have not been taken into account here. If model is used to support moderation on social media, we recommend that final judgments should always be left for human moderators.

<!-- ### Out-of-Scope Use *List how the model may foreseeably be misused and address what users ought not to do with the model.* --> <!-- ## Bias, Risks and Limitations *What are the known or foreseeable issues stemming from this model? You could also flag here known failure cases or weaknesses of the model.* --> <!-- ### Recommendations *What are recommendations with respect to the foreseeable issues? For example, filtering explicit content.* -->

Training Details

Training Set Metrics

Training Hyperparameters

• batch_size: (16, 16)
• num_epochs: (4, 4)
• max_steps: -1
• sampling_strategy: oversampling
• num_iterations: 6
• body_learning_rate: (2e-05, 1e-05)
• head_learning_rate: 0.01
• loss: CosineSimilarityLoss
• distance_metric: cosine_distance
• margin: 0.25
• end_to_end: False
• use_amp: False
• warmup_proportion: 0.1
• l2_weight: 0.01
• seed: 42
• evaluation_strategy: epoch
• eval_max_steps: -1
• load_best_model_at_end: False

Framework Versions

• Python: 3.11.9
• SetFit: 1.1.3
• Sentence Transformers: 3.2.0
• Transformers: 4.44.0
• PyTorch: 2.4.0+cu124
• Datasets: 2.21.0
• Tokenizers: 0.19.1

Citation

BibTeX

@article{paakki-implicit-indirect,
    doi = {https://doi.org/10.3384/nejlt.2000-1533.2025.5980},
    url = {https://nejlt.ep.liu.se/article/view/5980},
    author = {Paakki, Henna and Toivanen, Pihla and Kajava, Kaisla},
    title = {Implicit and Indirect: Detecting Face-threatening and Paired Actions in Asynchronous Online Conversations},
    publisher = {Northern European Journal of Language Technology (NEJLT)},
    volume= {11},
    number= {1},
    year = {2025}
}

<!-- ## Glossary *Clearly define terms in order to be accessible across audiences.* --> <!-- ## Model Card Authors *Lists the people who create the model card, providing recognition and accountability for the detailed work that goes into its construction.* --> <!-- ## Model Card Contact *Provides a way for people who have updates to the Model Card, suggestions, or questions, to contact the Model Card authors.* -->