metadata
license: mit
base_model: google/gemma-3-270m-it
tags:
- causal-reasoning
- semantic-loss
- fine-tuned
- transitivity
language:
- en
Gemma Transitivity Semantic V4
Fine-tuned Gemma 270M-IT for causal transitivity reasoning using dynamic semantic loss scheduling (λ: 0.05→0.30).
Performance
- Standard accuracy: 70.4% (vs 27.7% collapsed baseline)
- Adversarial accuracy: 69.8%
- Branching task: 97.9% (vs 1.96% baseline)
Key Innovation
Dynamic lambda scheduling prevents model collapse — a critical failure mode where standard fine-tuning causes models to predict constant "Yes"/"No" regardless of input.
Usage
from transformers import AutoTokenizer, AutoModelForCausalLM
model = AutoModelForCausalLM.from_pretrained("ludwigw/gemma-transitivity-semantic-v4")
tokenizer = AutoTokenizer.from_pretrained("ludwigw/gemma-transitivity-semantic-v4")
Citation
@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}
}