grimjim's picture
Update README.md
8ab69b6 verified
---
base_model: grimjim/kunoichi-lemon-royale-v3-32K-7B
library_name: transformers
quanted_by: grimjim
license: cc-by-nc-4.0
pipeline_tag: text-generation
---
# kunoichi-lemon-royale-v3-32K-7B
This is a merge of pre-trained Mistral 7B language models created using [mergekit](https://github.com/cg123/mergekit).
With this merger, we explore *merge densification*, a merge approach that attempts to transfer and adapt some benefits of denser models. A highly creative model, which itself was merged from multiple dense models, was merged in at very low weight in order to lightly modify the base model. The result was expected to improve variability in output without significantly impacting the coherence in the base model.
Tested with ChatML instruct templates, temperature 1.0, and minP 0.02. Practical context length should be at least 16K.
The additional model merge weight of 0.02 was deliberately chosen to be on par with the minP setting.
- Full weights: [grimjim/kunoichi-lemon-royale-v3-32K-7B](https://huggingface.co/grimjim/kunoichi-lemon-royale-v3-32K-7B)
- GGUF quants: [grimjim/kunoichi-lemon-royale-v3-32K-7B-GGUF](https://huggingface.co/grimjim/kunoichi-lemon-royale-v3-32K-7B-GGUF)
## Merge Details
### Merge Method
This model was merged using the [task arithmetic](https://arxiv.org/abs/2212.04089) merge method using [grimjim/kunoichi-lemon-royale-v2-32K-7B](https://huggingface.co/grimjim/kunoichi-lemon-royale-v2-32K-7B) as a base.
### Models Merged
The following model was also included in the merge:
* [grimjim/rogue-enchantress-32k-7B](https://huggingface.co/grimjim/rogue-enchantress-32k-7B)
### Configuration
The following YAML configuration was used to produce this model:
```yaml
base_model: grimjim/kunoichi-lemon-royale-v2-32K-7B
dtype: bfloat16
merge_method: task_arithmetic
slices:
- sources:
- layer_range: [0, 32]
model: grimjim/kunoichi-lemon-royale-v2-32K-7B
- layer_range: [0, 32]
model: grimjim/rogue-enchantress-32k-7B
parameters:
weight: 0.02
```