Standard fine-tuned Gemma 270M-IT on d-separation task without semantic loss. Exhibits model collapse — predicts near-constant "No" (7.6% F1). Provided for reproducibility and as a negative baseline.

Usage

python

from transformers import AutoTokenizer, AutoModelForCausalLM

model = AutoModelForCausalLM.from_pretrained("ludwigw/gemma-dseparation-baseline")
tokenizer = AutoTokenizer.from_pretrained("ludwigw/gemma-dseparation-baseline")

Citation

bibtex

@article{deshmukh2026semantic,
  title={On Semantic Loss Fine-Tuning Approach for Preventing Model Collapse in Causal Reasoning},
  author={Deshmukh, Pratik and Gupta, Atirek},
  journal={arXiv preprint arXiv:2605.05438},
  year={2026}
}