CLArg-group/ita-bsm-q-sem

ITA base score model: bsm-q-sem (BSM-Q (primed & sem))

A base score model (BSM) for inference-time argumentation (ITA), from the paper Neurosymbolic Learning for Inference-Time Argumentation. Given a claim and an argument supporting or attacking it, the BSM assigns the argument a base score in [0, 1], representing its intrinsic strength. ITA combines these base scores with the DF-QuAD gradual semantics to compute the strength of the claim and a True / False / Uncertain verdict.

It is referred to as BSM-Q (primed & sem) in the paper.

Model

• Backbone: Qwen/Qwen3-8B with a LoRA adapter (adapter/)
• Regression head: regression_head.safetensors, configured by model_config.json
• Training arguments: generated by Qwen3-8B for the TFU Training Claims (Data/Training/training_true_false_uncertain_with_arguments_qwen.json)
• Training: the regression head was first primed on base scores obtained by prompting the backbone, then the head and the adapter were trained end to end through the DF-QuAD semantics with the semantic loss.

Usage

The model is loaded with the code in the ITA repository. Download it into checkpoints/bsm-q-sem/, where the repository's configs expect it:

hf download <this repo> --local-dir checkpoints/bsm-q-sem

To score arguments from a CSV file with claim,is_support,argument columns:

python -m ita.base_score_model.infer_on_arguments \
    --checkpoint-dir checkpoints/bsm-q-sem --input-csv arguments.csv --output-csv scores.csv

See the repository README for training, evaluation and the other released BSMs.

License

The model weights are released under the MIT license. Use of the model is also subject to the licence of the base model (Qwen/Qwen3-8B) and to the terms of the data it was trained on, as described in the repository.