In an article it was relieved that 80000 Google Employees tested its GPT, which still makes errors in outputs.
Now there are some key points. These points can be used on any GPT testing, not necessarily of capacity of 80000. When someone makes testing which is so huge, 80000 is a huge number, then one must document the testing process as follows:
-The test cases, means here the latest GPT version test cases.
-What was the search string or prompt?
-What was the output produced by GPT version under test ?
-What was engineer testing rating for the output.
– Does the engineer call it a good response or ok response or a negative response
-Also ask for a rating in -10 to 10 range
– Any other tests done by engineers to be logged in the log book in same way.
This is a database and even possibly correct answer can be fed in this database.
– These test cases can help frame the right outputs when they are used in reinforcements learning especially when 80000 engineers test, this can generate huge data to give back to GPT to improve results.
-Many users can use log book at same time with synchronization being on for logging in
-Given each engineer must have tested at least 4-5 string search operations that means 80000*5= 400000 entries in testing log books can be made, these can be used in future GPT versions as well, to improve, to do reinforcement learning and even to test newer versions.
Hence, it’s not enough just to test GPT, if you test GPT, log the details. Reuse it to mold new GPT versions. These logging in can help in automated testing as well as to shape in the reinforcement learning in the GPT to make GPT new versions more human like, and more error free.
So the next levels would be evaluated and this molds GPT. This also automate testing. These testing data’s can be generated which are fully voluntary and which are open source. However, companies can pay to make their own datasets areas well to ensure their own quality and protocols.
Artificial Intelligence, Computer Science & Mathematics 