license: cc-by-4.0
language:
- en
tags:
- merge
- not-for-all-audiences
- nsfw
Top 1 RP Performer on MT-bench 🤪
Next Gen Silicon-Based RP Maid
WTF is This?
Silicon-Maid-7B is another model targeted at being both strong at RP and being a smart cookie that can follow character cards very well. As of right now, Silicon-Maid-7B outscores both of my previous 7B RP models in my RP benchmark and I have been impressed by this model's creativity. It is suitable for RP/ERP and general use.
It's built on xDAN-AI/xDAN-L1-Chat-RL-v1, a 7B model which scores unusually high on MT-Bench, and chargoddard/loyal-piano-m7, an Alpaca format 7B model with surprisingly creative outputs. I was excited to see this model for two main reasons:
- MT-Bench normally correlates well with real world model quality
- It was an Alpaca prompt model with high benches which meant I could try swapping out my Marcoroni frankenmerge used in my previous model.
MT-Bench Average Turn
model | score | size |
---|---|---|
gpt-4 | 8.99 | - |
xDAN-L1-Chat-RL-v1 | 8.24^1 | 7b |
Starling-7B | 8.09 | 7b |
Claude-2 | 8.06 | - |
Silicon-Maid | 7.96 | 7b |
Loyal-Macaroni-Maid | 7.95 | 7b |
gpt-3.5-turbo | 7.94 | 20b? |
Claude-1 | 7.90 | - |
OpenChat-3.5 | 7.81 | - |
vicuna-33b-v1.3 | 7.12 | 33b |
wizardlm-30b | 7.01 | 30b |
Llama-2-70b-chat | 6.86 | 70b |
^1 xDAN's testing placed it 8.35 - this number is from my independent MT-Bench run.
It's unclear to me if xDAN-L1-Chat-RL-v1 is overtly benchmaxxing but it seemed like a solid 7B from my limited testing (although nothing that screams 2nd best model behind GPT-4). Amusingly, the model lost a lot of Reasoning and Coding skills in the merger. This was a much greater MT-Bench dropoff than I expected, perhaps suggesting the Math/Reasoning ability in the original model was rather dense and susceptible to being lost to a DARE TIE merger?
Besides that, the merger is almost identical to the Loyal-Macaroni-Maid merger with a new base "smart cookie" model. If you liked any of my previous RP models, give this one a shot and let me know in the Community tab what you think!
The Sauce
models: # Top-Loyal-Bruins-Maid-DARE-7B
- model: mistralai/Mistral-7B-v0.1
# no parameters necessary for base model
- model: xDAN-AI/xDAN-L1-Chat-RL-v1
parameters:
weight: 0.4
density: 0.8
- model: chargoddard/loyal-piano-m7
parameters:
weight: 0.3
density: 0.8
- model: Undi95/Toppy-M-7B
parameters:
weight: 0.2
density: 0.4
- model: NeverSleep/Noromaid-7b-v0.2
parameters:
weight: 0.2
density: 0.4
- model: athirdpath/NSFW_DPO_vmgb-7b
parameters:
weight: 0.2
density: 0.4
merge_method: dare_ties
base_model: mistralai/Mistral-7B-v0.1
parameters:
int8_mask: true
dtype: bfloat16
For more information about why I use this merger, see the Loyal-Macaroni-Maid repo
Prompt Template (Alpaca)
I found the best SillyTavern results from using the Noromaid template but please try other templates! Let me know if you find anything good.
SillyTavern config files: Context, Instruct.
Additionally, here is my highly recommended Text Completion preset. You can tweak this by adjusting temperature up or dropping min p to boost creativity or raise min p to increase stability. You shouldn't need to touch anything else!
Below is an instruction that describes a task. Write a response that appropriately completes the request.
### Instruction:
{prompt}
### Response: