yhua219/EduRABSA_SLM_v1_SLERP_phi4mini_onnx

Licence and Copyright

• The "EduRABSA_SLM" LoRA adaptors and any merged models derived from them are licensed under the Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International License.

• Copyright (c) 2025 Authors of Data-Efficient Adaptation and a Novel Evaluation Method for Aspect-based Sentiment Analysis.

• The two pre-trained base models ("the original models") Phi4-mini-instruct and Qwen2.5-1.5B-Instruct were used for inference and training LoRA adaptors. No modifications were made to the original models. The original models’ licences and copyright notices (where provided) are included in the corresponding subdirectories of this repository.

The EduRABSA_SLM Model Family

The "EduRABSA_SLM" model family consists of fine-tuned multi-task small LLMs (SLMs) designed for resource-efficient opinion mining on education-domain reviews of courses, teaching staff, and universities (e.g. student course or teaching evaluations, and open-ended survey responses).

To cite the LoRA adaptors or merged models in this family:

@misc{hua2025dataefficientadaptationnovelevaluation,
      title={Data-Efficient Adaptation and a Novel Evaluation Method for Aspect-based Sentiment Analysis}, 
      author={Yan Cathy Hua and Paul Denny and Jörg Wicker and Katerina Taškova},
      year={2025},
      eprint={2511.03034},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2511.03034}, 
}

The EduRABSA_SLM multi-task models can perform opinion mining across the following fine-grained Aspect-based Sentiment Analysis (ABSA) tasks, extracting outputs for and across each review entry as illustrated in the image below:

• Opinion Extraction (OE)
• Aspect-Opinion Pair-Extraction (AOPE)
• Aspect-opinion Categorisation (AOC; ASC with opinion term)
• Aspect-(opinion)-Sentiment Triplet Extraction (ASTE)
• Aspect-(opinion-category)-Sentiment Quadruplet Extraction (ASQE)

Model Info

This is a CPU-only, ONNX format, quantised version of one of the "EduRABSA_SLM_v1_SLERP" models correspond to the following models described in the paper:

• EduRABSA_SLM_v1_SLERP_phi4mini = merged_Phi4_SLERP, base model Phi4-mini-instruct.

• EduRABSA_SLM_v1_SLERP_qw2.5-1.5B = merged_Qwen2.5_SLERP, base model Qwen2.5-1.5B-Instruct.

Each of these models was created via weight-merging (using the SLERP algorithm) two LoRA fine-tuned models on the same base model, and shares the multi-task capabilities described above. The EduRABSA_SLM LoRA adaptors were fine-tuned using the EduRABSA dataset.

The model follows its base model's hardware and environment requirements, unless the model name contains onnx (in which case, please refer to the onnx-model's README file).

Full details regarding the development and performance of the EduRABSA_SLM models are available in our paper Data-Efficient Adaptation and a Novel Evaluation Method for Aspect-based Sentiment Analysis.

Minimum Requirements

Please visit here for the platform and architecture supported by onnxruntime-genai. Note this model is CPU-only.

The package version numbers below are for reference in case you encounter issues during installation or model use.

# python==3.12

onnxruntime==1.23.2
onnxruntime-genai==0.10.0

How to Use

• Please see usage_example.ipynb for an example (code and prompt) of using the model for the ASQE task.

• For the prompts used for training the five education review ABSA tasks, please visit the project's GitHub repository.