Exploring and Evaluating Various Domain Adaptation Techniques
02/17/2025Domain adaptation involves modifying a model, originally trained on general data, to perform effectively in more specialized fields such as medicine, politics, law or instruction following.
This thesis will explore different adaptation approaches, like full finetuning or adapters, to finetune our pretrained language model LLäMmlein, for these specific domains.
Objectives:
- Create and preprocess domain datasets & benchmarks
- Explore and compare multiple adaptation techniques
- Evaluate effectiveness of different approaches
Requirements:
- Proficiency in Python coding, with a preference for experience in neural network and natural language processing libraries, like Pytorch and Huggingface
- For the duration of your thesis, you will have access to our computing cluster to train, execute, and evaluate your experiments
Betreuer: Julia Wunderle
