The-Face-Of-Goonery/HuginnV5.5-12.6B

(disclaimer, for new downloaders. please refer to the v56/ v5.6 version of Huginn, it is noticably improved even over 5.5, there is a new gguf file, and a zip file containing the fp16 weights available)

Huginn V5.5

Experimental frankenmerge using multiple 7B models using the Dare-ties method.

Including:

Part 1:

• https://huggingface.co/jondurbin/bagel-dpo-7b-v0.1
• https://huggingface.co/maywell/Synatra-7B-v0.3-RP

Part 2:

• https://huggingface.co/mlabonne/NeuralBeagle14-7B
• https://huggingface.co/openaccess-ai-collective/DPOpenHermes-7B-v2

Part 3:

merged part 1 and part 2 together

Part 4:

then took the first 26 layers of https://huggingface.co/FelixChao/WestSeverus-7B-DPO-v2 and added them before the 32 layers of part 3 to make the final model

Prompting and scope:

seems to work well with alpaca for instructions, and chatML format for just normal conversation.

scores like just under 73 points on the leaderboard, way higher than any huginn model before, by a factor of around 10 points.

Huginn primarily excells at conversational tasks, and creative tasks, being capable at story writing, roleplaying, even helping writers with creative tasks, (Huginn is capable of coming up with creative ideas better than most other models)