File size: 3,297 Bytes
f0c08d0 9c7efad f0c08d0 dba74f0 f0c08d0 4076066 f0c08d0 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 |
---
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
```
|