Details
-
Improvement
-
Resolution: Unresolved
-
Minor
-
None
-
None
-
None
-
3
-
9223372036854775807
Description
Since we have a DGX sitting mostly unused (gpu-wise), I thought it would be an interesting idea to run it on various documentation we have to highlight problems. This resulted in various useful actionable data and Andreas suggested to do the same on the Lustre manual.
This also seems to have returned useful data, I am going to post the structured results here hoping somebody could take a look and push patches as needed.
Do not take the suggestions as gospel, but in general it highlights somewhat problematic areas and even if you use a different way to fix it - that's still beneficial. In particular suggestions like "use file system or filesystem consistently" are good, but unlike what the model proposes, we actually want to use filesystem as a single word.
The results from my manpage runs suffered from some hallucinations including some not really there spelling mistakes, so don't be surprised if you encounter something like this, but this is a much larger model result, so should be better at not doing this. In fact I did some spotchecking and it looks like all spelling errors below for the most part are not there, may be the prompt needs to be improved to say that it's ok for there not to be any spelling errors?
This is the prompt I used (model llama 3.1 405b quantified to 4 bits from 16 so it actually fits):
prompt = """You are an expert proofreader and your job is to review Lustre filesystem documentation.
Read the below section in XML format for spelling and grammar errors, inconsistencies and poor wording:
```
%s
```Proofread the above document and provide your feedback.
Be specific and short, recommend suggested fixes for misspellings and grammar errors.
Limit every suggestion to 50 words or less.
Explain inconsistencies.
Respond using JSON like this:
{"spelling":[{"original":..., "fix":...}],"grammar":[{"original":...,"fix":...}],"inconsistency":[{"original":...,"explanation":...}],"wording":[{"original":...,"suggestion":...}]}
"""
As you can see, the results are JSON split into different categories, hopefully even if you don't know what json is too deeply, the data is useful.