Datasets:
Releasing redacted training data too?
Hello! My name is Rahul and I'm a PhD student with Satra Ghosh at MIT. Excited to explore the things you open-sourced.
I was curious whether you were going to release a similar parquet file with the train data as well?
Thank you!
Rahul
Hi Rahul!
We hadn't planned to because the model scores on the training data are heavily biased by the training process. The main intention behind releasing the redacted data was for various types of threshold calibration, which are best done on data the model wasn't trained on. Since we can't release the audio data itself as part of the redacted dataset, it didn't seem like there would be much use to releasing a redacted version of the training set. But if you could describe your use case I can look into whether it's still something we could do.
Noah
Hi Noah!
I'm also a researcher in the voice biomarker/mental health space. My main goal is to use your DAM model as a comparison against the models I ran by feeding it my audio data. A part of my post-hoc analysis on my data was looking at the symptom distribution of the various PHQ-9 and GAD-7 scores as a way to understand what the model might have been doing. As such, I have two use-cases:
- I wanted to compare the distribution of sum scores and symptoms in my dataset to yours as a first EDA step.
- I also want to compare model performance (e.g. R2 scores, MAE)
For 2), I don't need the training data for the reasons you say. For 1), I would love the demographics (but not necessarily the predictions) just to have a much larger population to explore symptoms/demographics in. Of course the 15k samples in train/validate are ok for this, but a larger sample is always useful for that.
Cheers,
Rahul