Some suggestions

[trangdata]

Here are some of my suggestions after the first pass. Please feel free to accept/edit/decline as you see fit. Main points:

• Reduce frequency of "allow"
• When we start a sentence with "this", I try to specify what "this" is
• When reporting P value, I think it's clear that BioChatter's prompt engine yields significantly higher accuracy based on confidence intervals. I don't think we need to report the exact value for P. https://www.ncbi.nlm.nih.gov/pmc/articles/PMC8005799/#:~:text=If%20p%20values%20are%20reported,and%20for%20p%20values%20greater

[trangdata]

One comment I have is wrt Fig. 3 biochatter_benchmark: it's a little difficult for me to read/interpret this figure, e.g. I can't find values for chatglm3. I know it's challenging because we don't have the full grid of models x sizes x quantisation levels, but I wonder if jittering or faceting would help somehow.

Also, in the main text, we haven't discussed model "size", and given that there is not really an association between size and performance, maybe we don't need to visualize this dimension? (We can include a table of all the values and sizes in Supplement) And if so, I wonder if a graph of Mean Accuracy vs. Model would be a little clearer, and the point size would refer to the quantisation level instead. Just a thought.

[slobentanzer]

Good points, thanks a lot for the feedback! Will go over and merge.