| isarantopoulos | |
| Apr 3 2023, 4:27 PM |
| F36941539: Screenshot 2023-04-06 at 5.41.19 PM.png | |
| Apr 6 2023, 2:45 PM |
| F36941535: Screenshot 2023-04-05 at 7.11.17 PM.png | |
| Apr 6 2023, 2:45 PM |
| F36941537: old_gpt.png | |
| Apr 6 2023, 2:45 PM |
With the release of GPT-4 we want to experiment with this version which will allow us to provide more information from Wikipedia into the prompt.
The newer models allow n input of 8k-32k tokens (as opposed to gpt-3.5 where max of 4k tokens were allowed.
The intuition is that this will allow us to provide more informative and richer answers as the model will have more Wikipedia content in its context.
| Title | Reference | Author | Source Branch | Dest Branch | |
|---|---|---|---|---|---|
| Update WikiGPT with GPT-4 | toolforge-repos/wiki-gpt!16 | isaranto | add-gpt-4 | main |
| Status | Subtype | Assigned | Task | ||
|---|---|---|---|---|---|
| Resolved | None | T328494 WikiGPT Experiment | |||
| Resolved | isarantopoulos | T333854 GPT-4 in WikiGPT |
Created a merge request that does thes following:
I have synced this branch with the application at wiki-gpt-ml.toolforge.org for now.
According to OpenAI's release report “GPT-4 scores 40% higher than our latest GPT-3.5 on our internal adversarial factuality evaluations.
Some preliminary tests demonstrate that this is valid in our scenario. The tests were run before increasing the amount of text we pass to the model (same num of articles and paragraphs) and we expect the results to be even better with the aforementioned increase in information the model is exposed to.
The below example is one that the previous version was failing.
Old version:
