Detailed description of model training/evaluation (publication/thesis?)
Hi Nizar,
Thanks for sharing this model!
I'm planning on evaluating this model in some dialect ASR research of my own, and would love to be able to read up on more details regarding your fine-tuning and evaluation, if possible. From the other comments, I gather that this model was part of your Master's thesis. Are you considering publishing a paper on this project that I can keep an eye out for, and/or would you be happy to share your thesis (publicly or via email)?
Best,
Verena
(verena DOT blaschke AT cis.lmu.de)
Hello Verena,
Thank you very much for your interest in the model. Actually, I don't know what more details I could provide on the model. All the information is already available on this repo: base model, hardware, training time, fine-tuning technology (QLoRa) with its hyperparameters, train/test split, evaluation scores and comparison with other models. Tell me if I'm missing something, I'd be eager to help!
That said, as I could've seen in the ASR literature, we can't really publish something for this model, as long as we don't have human evaluation or other stuff to put in a potential paper. Also, my studies are not about NLP, it's a neuroscience master. I simply needed a better model for swiss-german so I did that as a side-project, whose, in fine, is the one that got the most attention.
My thesis will certainly be put online once it's done, but my focus is on computational neuroscience of language, so nothing that would interest pure-NLP labs I guess.
Best,
Nizar Michaud