Well done!
Well done. I've been following your work since Wraith; your fine-tunes are quite unique and well-executed. While at times they feel a bit too leaning towards safety, though this is simply from vibe testing. Realistically, your models are likely on par with base Gemma in terms of safety, with the heretic being an easy solution if that ever becomes an actual hindrance.
Scout, in its own way, was a charming tune; it was challenging to assess its practical utility due to the size constraints of the models, but an intriguing, novel approach nonetheless.
In an era where fine-tunes are tapering off as base models become increasingly overfit and benchmaxxed, your work has been a refreshing change of pace.
Keep up the excellent work, and thank you for sharing your innovations with the community!
Also, I am unsure if you are affiliated with this account. But, figured I'd at least bring it to your attention: https://huggingface.co/Cheeeeeeeeky/affine-homonculus
Wow, what a nice post! Definitely the highlight of my day, thank you so much! Your observations on my fine-tunes are actually really sharp - I'm not sure anyone has made those connections before (at least not to me 😂)
I'm really glad you like the Entities family - it's a fun project that's overdue for an update. I have a few ideas floating around for new Entities, so I'll have to jump on that! (If you ever have ideas for entities, feel free to send them my way - I'll cite you in the model card!)
Thank you for the note on safety! I've been playing around with safety/refusals a bit so that feedback is super helpful! I'll run some tests over here and see if I can make any improvements in that area. I've seen a few heretic iterations of my models floating around the hub, so I've kinda just left those to the community. I run everything myself over here so I just haven't had time to explore it.
Thanks for the heads up on the duplicate repo - this is the first I've heard of it. Not affiliated with me, but they're using the original model card and weights, so it's technically within open source norms I suppose. Annoying, but part of the territory 🤷♂️ Thanks for bringing it to my attention!