miqurelian-120b / README.md
jspr's picture
Update README.md
4b27fc1 verified
---
base_model:
- 152334H/miqu-1-70b-sf
- grimulkan/aurelian-v0.5-70b-rope8-32K-fp16
library_name: transformers
tags:
- mergekit
- merge
---
# miqurelian-120b
This is a 120b merge created by interleaving layers of [miqu-1-70b-sf](https://huggingface.co./152334H/miqu-1-70b-sf) with [Aurelian](https://huggingface.co./grimulkan/aurelian-v0.5-70b-rope8-32K-fp16), a creative writing model, using [mergekit](https://github.com/cg123/mergekit). It performs approximtely SOTA for long-context creative writing tasks that require strong semantic coherence.
## Model Details
- Max Context: 32768 tokens
- Layers: 140
### Prompt template
```
<s>[INST] {prompt} [/INST]
```
### Merge Method
This model was merged using the [linear](https://arxiv.org/abs/2203.05482) merge method.
### Models Merged
The following models were included in the merge:
- [152334H/miqu-1-70b-sf](https://huggingface.co./152334H/miqu-1-70b-sf)
- [grimulkan/aurelian-v0.5-70b-rope8-32K-fp16](https://huggingface.co./grimulkan/aurelian-v0.5-70b-rope8-32K-fp16)
### Configuration
The following YAML configuration was used to produce this model:
<details><summary>mergekit_config.yml</summary>
```yaml
merge_method: linear
parameters:
weight: 1.0
slices:
- sources:
- model: 152334H/miqu-1-70b-sf
layer_range: [0, 1]
- model: grimulkan/aurelian-v0.5-70b-rope8-32K-fp16
layer_range: [0, 1]
parameters:
weight: 0
- sources:
- model: 152334H/miqu-1-70b-sf
layer_range: [1, 20]
- sources:
- model: grimulkan/aurelian-v0.5-70b-rope8-32K-fp16
layer_range: [10, 30]
- sources:
- model: 152334H/miqu-1-70b-sf
layer_range: [20, 40]
- sources:
- model: grimulkan/aurelian-v0.5-70b-rope8-32K-fp16
layer_range: [30, 50]
- sources:
- model: 152334H/miqu-1-70b-sf
layer_range: [40, 60]
- sources:
- model: grimulkan/aurelian-v0.5-70b-rope8-32K-fp16
layer_range: [50, 70]
- sources:
- model: 152334H/miqu-1-70b-sf
layer_range: [60, 79]
- sources:
- model: 152334H/miqu-1-70b-sf
layer_range: [79, 80]
- model: grimulkan/aurelian-v0.5-70b-rope8-32K-fp16
layer_range: [79, 80]
parameters:
weight: 0
dtype: float16
tokenizer_source: model:152334H/miqu-1-70b-sf
```
</details>