Open james777b opened 4 months ago
Thanks for the report @james777b. The compression prompts are probably not "strong" enough to work well with smaller models, and should be revised.
It's impressive how the model hallucinated that code and also the username. Definitely it's not in the prompting, I wonder how they made it up.
Description
After an ongoing conversation (more than 4K tokens) with multiple models (llama2, codellama, tinyllama) via ollama, I compressed the conversation to the 'detailed' level. It worked fine afaict. Then I compressed again to 'brief.'
The text it rendered was: """ \begin{code} df = pd.read_csv('C:/Users/Administrator/Desktop/test.csv', skiprows=range(0, 4500)) \end{code} \end{code} """
That is four lines verbatim, even the duplicated end.
When I accepted that change to replace text, it did in fact replace the conversation text with message.
Also, that directory... isn't mine. I am running the app via a similar location; it's was in my username. And, I'm on Windows.
Device and browser
Windows 10; Firefox.
Screenshots and more
No response
Willingness to Contribute