It was mentioned that we are in the open ai dark winter; so I thought I would make myself a nice winter garden.
An experiment
I've merged four partitions successfully in the past, so lets go for 9! I started with:
Mistral-7B-v0.1
and merged in
OmniBeagleSquaredMBX-v3-7B
ZySec-7B-v1
Omningotex-7b-slerp
Erosumika-7B
LemonadeRP-4.5.3
Thespis-Krangled-7b
pastiche-crown-clown-7b-dare
Snorkel-Mistral-PairRM-DPO
multi_verse_model
9-partition merge
All of the layers were partitioned in to 9 random bins. Alternating models were slerped at [0...1], and [1...0] gradients; except attention, which was slerped at 0.03.
This means that the model is still predominantly ordered around base mistral - including half of the input and output layers, and 28% of attention.
Other
Includes fast tokenizer.
Chat Template
I put a conversational chat template, which takes "name", "to" (optional), and "content" as the turns. It is designed to follow a transcript style chat which is used by some of the models. This type of use-case is best done by outlining a scene and creating a character card.
### {% title %}
{% metadata %}
USER: Hello
ASSISTANT: Hi, how are you?
It leans to being a coder when given an ### Instruction, follows <s>[INST][/INST], and likes <|user|>, <|assistant|> as well.
A quite cheery and intelligent model. Very good with science and math, but still capable of a decent amount of creativity for a 7b model.
