Daredevil
I've been keeping an eye on Daredevil as well. Great minds think alike XD
mlabonne just released a DPO version: https://huggingface.co/mlabonne/NeuralDaredevil-8B-abliterated
I'll test Mahou 1.3 when I have time, that merge looks very appealing.
Ngl, it's hilarious seeing my models in the family history of the currently highest scoring MMLU llama 3 model. Mad respect to Maxime though, I'd be totally lost in this space without him.
Yeah I'll warn you though that the Mahou 1.3 finetune seems to be pretty much broken on L3. I'll probably replace or patch the model when I get a chance later. Still trying to figure out ORPO and handling ChatML tokens properly...
Ngl, it's hilarious seeing my models in the family history of the currently highest scoring MMLU llama 3 model. Mad respect to Maxime though, I'd be totally lost in this space without him.
Yeah I'll warn you though that the Mahou 1.3 finetune seems to be pretty much broken on L3. I'll probably replace or patch the model when I get a chance later. Still trying to figure out ORPO and handling ChatML tokens properly...
Already downloaded it T_T Alright, I'll wait until you fix it to redownload. Yeah, Mahou models perform surprisingly high in the LLM Open Leaderboard. It's hard for a L3 model to break the 70 average mark.
To be frank, I always use Llama 3 system instead of ChatML, but replacing user and assistant with {{user}} and {{char}}. Works like a charm.
Hmmm... is Mahou 1.3a ok to download? I'll wait until you green-light it. xD
You can try it but I think I'm going to do a retrain and rerelease for 1.3. And distribute an instruct template with it.
Ok, I'll wait then. ^_^' I trust in your judgement regarding this matter.
ChatML should be setup properly for this 1.3: https://huggingface.co/flammenai/Mahou-1.3-llama3-8B
It produces garbage at the end of messages sometimes but it seems to work decently otherwise. We'll see how GGUF quants go 😅