Questions?
Hi!
I am curious what is your feeling of this model? How is it in regards to positivity bias? Also, what is the context limit, before it starts to degrade?
I only did a quick run on a few cards so I don't really know much. Will test it more tomorrow/weekend
If it still has slop it most likely retained the positivity bias. model_stock also tends to make models more prone to refusals. For context lenght, Llama 3.3 is believed to be able to push beyond 32k while remaining coherent, however the fine-tuning and merging process may have diminished this.
Thanks for the answers, it helps. I can run as a test on a few cards which are more complex in scenario. Refusals are part of the scenarios, so we will see how deep they go. I can give you some feedback if you would like after I run the tests.
If it stays coherent at 32k, or 28k even, then I am really happy :)
Yeah, it's always interesting to read the feedback
After using it for a little longer, here are some of my own observations:
Good points:
- Dialogue flow. Model easily picks up when I'm joking, being serious or flirting without me stating it. The character portrayal was also nice and appropriate, thought I mostly chatted with funnier and lighter ones. I don't know how well the model works with more grim types
- Creativity. Swipes lead to different outcomes, and I really dig it as I love just sitting and pocking until the "ideal" output appears. The model also throws in some interesting phrases from time to time, though I would like it to do this more often.
- Сensorship. Actually, I didn't encounter any refusals at any point, even when things got gory. The model didn't describe much though, and that leads to the downside of this thing
The cons: - Everything else feels like generic LLM stuff. The slop, the summaries, the unwillingness to go deep into the NSFW scenes (or any scenes for that matter). As I was writing this, I remembered that I didn't try to stir it up with an author's note. Might be worth poking around, but probably won't help much
It also seems to benefit from having a ChatML as an instruct template, so with that my current setup looks like this: https://files.catbox.moe/xvic3m.json