Proposal: Benchmarking on different PEFT techniques

I recently read this Hugging Face blog: "Beyond LoRA: Can You Beat the Most Popular Fine-Tuning Technique?" The blog concludes that, depending on the downstream task, PEFT techniques other than LoRA can outperform it. Inspired by this benchmark, I'd like to contribute by evaluating different PEFT methods on Table Recognition and OCR tasks using datasets available on Hugging Face.

I plan to run a comprehensive set of PEFT experiments on OCR-focused vision-language models such as:

• zai-org/GLM-OCR
• PaddlePaddle/PaddleOCR-VL-1.6
• and other relevant OCR/Table Recognition models.

Example dataset: apoidea/fintabnet-html

The goal is to benchmark different PEFT methods across these tasks and compare their effectiveness in terms of accuracy, efficiency, and resource usage.

If this sounds like a valuable addition to the benchmark, I'd be happy to take ownership of this task. Please assign it to me if you think it would be a good fit.