Open VolhaBazhkova opened 10 months ago
Author is @adubovik
We need to improve Markdown recognition in the model output. Currently, it's assumed that the entire output of the model is always markdown.
This leads to problems when the model
See examples attached.
Check how it works in ChatGPT - it recognizes Markdown-like markup locally, it doesn't apply it to the whole model output.
[Aliaksei Vavilau] Requires investigation. Should we use heuristic? If yes, then what is the logic?
As far as I see the 1st example with $ sign is not reproduced anymore; but the 2nd one is still reproduced
epam_aidial-_development_chat_conversation_4-9.json
Author is @adubovik
We need to improve Markdown recognition in the model output. Currently, it's assumed that the entire output of the model is always markdown.
This leads to problems when the model
See examples attached.
Check how it works in ChatGPT - it recognizes Markdown-like markup locally, it doesn't apply it to the whole model output.
[Aliaksei Vavilau] Requires investigation. Should we use heuristic? If yes, then what is the logic?