|
--- |
|
base_model: |
|
- cognitivecomputations/dolphin-2.9-llama3-8b |
|
- meta-llama/Meta-Llama-3-8B-Instruct |
|
library_name: transformers |
|
tags: |
|
- mergekit |
|
- merge |
|
- llama |
|
- llama3 |
|
license: other |
|
license_name: llama3 |
|
license_link: LICENSE |
|
--- |
|
|
|
# Model Details |
|
|
|
Uses ChatML but Alpaca probably works as well. |
|
|
|
[Roleplaying presets for SillyTavern](https://huggingface.co/Virt-io/SillyTavern-Presets) |
|
|
|
Configs copied from: |
|
- [chargoddard/mistral-11b-slimorca](https://huggingface.co/chargoddard/mistral-11b-slimorca) |
|
- [Replete-AI/Llama-3-11.5B-V2](https://huggingface.co/Replete-AI/Llama-3-11.5B-V2) |
|
- [abacusai/TheProfessor-155b](https://huggingface.co/abacusai/TheProfessor-155b) |
|
|
|
|
|
A try at a larger llama3 model. |
|
|
|
Using [cognitivecomputations/dolphin-2.9-llama3-8b](cognitivecomputations/dolphin-2.9-llama3-8b) for an uncensored base and [meta-llama/Meta-Llama-3-8B-Instruct](https://huggingface.co/meta-llama/Meta-Llama-3-8B-Instruct) as the duplicated layers as I really like its instructions following abilities. Hoping that it will be smarter and less censored. |
|
|
|
--- |
|
# llama3-11.5B |
|
|
|
This is a merge of pre-trained language models created using [mergekit](https://github.com/cg123/mergekit). |
|
|
|
## Merge Details |
|
### 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: |
|
* [cognitivecomputations/dolphin-2.9-llama3-8b](https://huggingface.co/cognitivecomputations/dolphin-2.9-llama3-8b) |
|
* [NousResearch/Meta-Llama-3-8B-Instruct](https://huggingface.co/NousResearch/Meta-Llama-3-8B-Instruct) |
|
|
|
### Configuration |
|
|
|
The following YAML configuration was used to produce this model: |
|
|
|
```yaml |
|
merge_method: linear # use linear so we can include multiple models, albeit at a zero weight |
|
parameters: |
|
weight: 1.0 # weight everything as 1 unless specified otherwise - linear with one model weighted at 1 is a no-op like passthrough |
|
slices: |
|
- sources: |
|
- model: cognitivecomputations/dolphin-2.9-llama3-8b # embed_tokens comes along with the ride with whatever is the first layer |
|
layer_range: [0, 1] |
|
- model: NousResearch/Meta-Llama-3-8B-Instruct # add dummy second model with 0 weight so tokenizer-based merge routine is invoked for embed_tokens |
|
layer_range: [0, 1] |
|
parameters: |
|
weight: 0 |
|
- sources: |
|
- model: cognitivecomputations/dolphin-2.9-llama3-8b |
|
layer_range: [1, 24] |
|
- sources: |
|
- model: NousResearch/Meta-Llama-3-8B-Instruct |
|
layer_range: [8, 24] |
|
parameters: |
|
scale: |
|
- filter: o_proj |
|
value: 0.0 |
|
- filter: down_proj |
|
value: 0.0 |
|
- value: 1.0 |
|
- sources: |
|
- model: cognitivecomputations/dolphin-2.9-llama3-8b |
|
layer_range: [24, 31] |
|
- sources: # same as above, but for lm_head with the last layer |
|
- model: cognitivecomputations/dolphin-2.9-llama3-8b |
|
layer_range: [31, 32] |
|
- model: NousResearch/Meta-Llama-3-8B-Instruct |
|
layer_range: [31, 32] |
|
parameters: |
|
weight: 0 |
|
dtype: bfloat16 |
|
tokenizer_source: model:cognitivecomputations/dolphin-2.9-llama3-8b # keep exact tokenizer used by dolphin - or you could use `union` if you add all of the input models to the first/last slice |
|
|
|
``` |
|
|
