CASE – Condition-Aware Sentence Embeddings for Conditional Semantic Textual Similarity Measurement [EACL2026]

Code Structure

File	Description
generate_embeddings.py	Step 1 of CASE: Creates initial unsupervised condition-aware embeddings from the LLM encoders and saves them to disk.
CASE_performance.py	Step 2 of CASE: Trains the supervised FFN on the saved embeddings and evaluates performance on the C-STS benchmark.
lora_finetune.py	Fine-tunes the base LLM encoders on the C-STS task using LoRA.
load_finetuned_llm.py	Loads the saved LoRA adapters to generate conditional embeddings from the fine-tuned model.
llm_encoders.py	Helper class containing the LLM_EMBEDDER wrapper for various LLM architectures.

Data Preparison

We use the C-STS-Reannotated dataset. Please download the training and validation files (csts_train_reannotated.csv and csts_validation_reannotated.csv) from Hugging Face repository: https://huggingface.co/datasets/LivNLP/C-STS-Reannotated

Paper Link

For more technical details, please refer to our full paper:
CASE -- Condition-Aware Sentence Embeddings for Conditional Semantic Textual Similarity Measurement

Citation

If you find this work or the code useful, please cite our paper:

@misc{zhang2025caseconditionawaresentence,
      title={CASE -- Condition-Aware Sentence Embeddings for Conditional Semantic Textual Similarity Measurement}, 
      author={Gaifan Zhang and Yi Zhou and Danushka Bollegala},
      year={2025},
      eprint={2503.17279},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2503.17279}, 
}