Nice model
#4
by
cmp-nct
- opened
Overall I believe this is better than ShareGPT (even the larger variant) so at first glance it appears to be the best llava 1.5 type I've seen.
One significant downside I noticed is that it is not good in following instructions and it is prone to repeating words.
So for example asking it to describe TWO things, it typically will respond with one thing only.
Or to ask it to list things, it will start listing and then the last one will be repeated.
That indicates an issue with fine tuning, the repeating problem could be hackily solved with sampling but the instruction issue would need a more diverse type of input.
ShareGPT is able to respond to 2 or more questions in one go