### This the config.yml for ABC_Books/test001 ### | |
models: | |
### Models that contribute a large 128K context window ### | |
- model: CallComply/zephyr-7b-beta-128k | |
parameters: | |
weight: 0.1154 | |
density: 0.9 | |
- model: Nitral-Archive/HerculeanSea-7b-128k | |
parameters: | |
weight: 0.1154 | |
density: 0.9 | |
- model: NousResearch/Yarn-Mistral-7b-128k | |
parameters: | |
weight: 0.1154 | |
density: 0.9 | |
### Models with finetune training on occult knownledge ### | |
- model: teknium/llama-deus-7b-v3-lora-merged | |
parameters: | |
weight: 0.0769 | |
density: 0.9 | |
- model: teknium/Hermes-Trismegistus-Mistral-7B | |
parameters: | |
weight: 0.0769 | |
density: 0.9 | |
- model: alexandrabenamar/Mistral-7B-Instruct-v0.2-Magic | |
parameters: | |
weight: 0.0769 | |
density: 0.9 | |
- model: tarotscientist/llama-2-7b-tarotreader | |
parameters: | |
weight: 0.0769 | |
density: 0.9 | |
- model: teknium/Mistral-Trismegistus-7B | |
parameters: | |
weight: 0.0769 | |
density: 0.9 | |
### Talkative model with a large context window ### | |
- model: Norquinal/Mistral-7B-storywriter | |
parameters: | |
weight: 0.0769 | |
density: 0.9 | |
### Models with finetune training to be uncensored use some crass diction ### | |
- model: Undi95/BigL-7B | |
parameters: | |
weight: 0.0384 | |
density: 0.9 | |
- model: Undi95/LewdMistral-7B-0.2 | |
parameters: | |
weight: 0.0385 | |
density: 0.9 | |
- model: Undi95/MistRP-Dolphin-7B | |
parameters: | |
weight: 0.0385 | |
density: 0.9 | |
- model: Undi95/Mistral-ClaudeLimaRP-v3-7B | |
parameters: | |
weight: 0.0385 | |
density: 0.9 | |
- model: Undi95/Toppy-M-7B | |
parameters: | |
weight: 0.0385 | |
density: 0.9 | |
### The use of DARES has been shown to “Densify” standard model lending to a more robust model when paired with a high “density:” numbers ### | |
merge_method: dare_ties | |
### This model best exemplifies the closest match to all of the features needed in the final model ### | |
base_model: NousResearch/Yarn-Mistral-7b-128k | |
parameters: | |
### When “densifing” models the model size tends to grow without normalize | |
normalize: true | |
int8_mask: true | |
dtype: float16 | |