Closed barun-saha closed 2 weeks ago
Mistral-7B-Instruct-v0.2 has been added back. This is in addition to Mistral-Nemo-Instruct-2407.
The former is typically faster but usually generates a shorter response (i.e., less content). The latter usually generates relatively detailed content.
A few other LLMs were tried out. However, they were largely found unsatisfactory for various reasons, e.g., their instruction-following capability to generate JSON output in the desired format was often sub-optimal.
Make Mistral v2 available as the only (or alternative) LLM for content generation.
The switch to Mistral Nemo was made to make use of larger context window, especially with the icons list on mind. The need is planned to be revisited (#48).
This issue should be addressed only after #48 is closed.