Magnum v4
#9
by
BigHuggyD
- opened
Wow, significant improvement to 'smarts' while keeping the same creative score mostly. Looks like I will need to check it out.
They've trained on more GPT data compared to v2, so don't be too optimistic if you hate it.
They've trained on more GPT data compared to v2, so don't be too optimistic if you hate it.
Mnnn, yes.. I am literally getting output with 'chatgpt:' as a prefix... ahh well... back to Behemoth
And again, same mistake as the last time: they didn't remove the refusals, making v4 even worse than v2 on UGI. When will they learn to filter the training data? There are plenty of <8b models which can be used for simple sentiment analysis, they don't even have to hire people.