Model
- Backbone: google/gemma-4-E4B-it with a LoRA adapter
(rank 8, α = 32, dropout 0.1, targeting the query, key, value and output projections)
- Base score model used in the reward: BSM-Q (primed & sem & rank & con), trained on Qwen3-8B-generated arguments with the semantic loss plus the ranking and consistency auxiliaries.
It is frozen during GRPO; only the generator's LoRA parameters are updated.
- Semantics: DF-QuAD, with the claim's own base score fixed at 0.5
- Verdict thresholds used for the reward: fixed, α = 1/3 and β = 2/3
Training
The generator was trained with GRPO (TRL) on the TFU Training
Claims (1,000 claims labelled True / False / Uncertain). Each sampled completion is parsed into
supporting and attacking arguments, scored by the BSM and evaluated with DF-QuAD; the resulting
claim strength is converted into the argumentation reward rα,β defined in Section 6 of the
paper, which is maximal for any strength that yields the correct verdict and decreases linearly with
distance from the relevant threshold. Unparseable completions receive a fixed reward of −1000.
Learning rate 1e-5, 2 epochs, 2 generations per prompt, 450 optimisation steps, seed 0.
Usage
The model is loaded with the code in the ITA repository.
Download it into checkpoints/ita-gemma4-grpo-args-bsm-q-sem-rank-con/, where the repository's configs expect it:
hf download CLArg-group/ita-gemma4-grpo-args-bsm-q-sem-rank-con --local-dir checkpoints/ita-gemma4-grpo-args-bsm-q-sem-rank-con
It can also be loaded directly as a PEFT adapter and prompted with the argument generation prompt
given in Appendix A.1 of the paper:
from peft import PeftModelfrom transformers import AutoModelForCausalLM, AutoTokenizer base = AutoModelForCausalLM.from_pretrained("google/gemma-4-E4B-it", dtype="bfloat16", device_map="auto")model = PeftModel.from_pretrained(base, "CLArg-group/ita-gemma4-grpo-args-bsm-q-sem-rank-con")tokenizer = AutoTokenizer.from_pretrained("CLArg-group/ita-gemma4-grpo-args-bsm-q-sem-rank-con")
See the repository README for training, evaluation and the other released models.
License
The model weights are released under the MIT license. Use of the model is also subject to the
licence of the base model (google/gemma-4-E4B-it) and to the terms of the data it was trained on, as described in
the repository.