|
--- |
|
license: mit |
|
language: |
|
- en |
|
--- |
|
|
|
Imatrix compressions of FP Merge of "D_AU-Orac-13B-Tiefighter-slerp". |
|
|
|
"Imatrix Plus" is an upgraded form of Imatrix which using full precision for specific parts of the compression. |
|
As a result all compressions will be slightly larger in size than standard 13B compressions. |
|
|
|
This method results in a higher quality model, especially at lower compressions. |
|
This method is applied across all compressions from IQ1 to Q8. |
|
|
|
Even IQ1_S - the most compressed verison - works well, however IQ4/Q4 are suggested as minimums for quality. |
|
Highest quality will be Q6/Q8. |
|
|
|
How big a difference is this merge? |
|
|
|
Orginal Tiefighter IQ1_S (with imatrix enhancements) tested at a perplexity of: |
|
PPL = 17.2589 +/- 0.12466* |
|
|
|
Tiefighter Orca 2 IQ1_S (with imatrix enhancements) tested at a perplexity of: |
|
PPL = 12.6985 +/- 0.09106* |
|
|
|
Note that LOWER perplexity is better. |
|
|
|
* Tested using llamacpp, perplexity.exe with wiki.raw. |
|
|
|
In addition the Imatrix file used to "fix" the compressed files post compression resulted in |
|
over 2 whole points lower perplexity at IQ1_S vs some of the other "Imatrix" files currently in use. |
|
|
|
Orginal Tiefighter IQ1_S (with imatrix enhancements) tested with a different "Imatrix" repair file at a perplexity of: |
|
PPL = 19.6355 +/- 0.14435 |
|
|
|
Likewise the merge itself affected perplexity too. |
|
|
|
This merge was an experiment to test already established Roleplay, Fiction and Story |
|
generation of "Tiefighter" with a some of "Orca 2"'s qualities. |
|
|
|
Additional merge experiements are in progress. |
|
|
|
For Imatrix plus this was a test of high precision in specific areas of the model leading to a slightly larger compressed file. |
|
In addition the Imatrix process itself used a larger "calibration" file than standard to further enhance quality. |
|
|
|
The process added appoximately 310 MB to each compressed file. |
|
|
|
A blank or standard Alpaca Template for text generation will work. |
|
Currently "CHATML" is untested. |
|
|
|
Context length: 4096. |
|
|
|
Please see the orginal model card for specific details of use, additional credits and tips: |
|
|
|
[KoboldAI/LLaMA2-13B-Tiefighter](https://huggingface.co/KoboldAI/LLaMA2-13B-Tiefighter) |
|
|
|
# merge |
|
|
|
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 SLERP merge method. |
|
|
|
### Models Merged |
|
|
|
The following models were included in the merge: |
|
* [microsoft/Orca-2-13b](https://huggingface.co/microsoft/Orca-2-13b) |
|
* [KoboldAI/LLaMA2-13B-Tiefighter](https://huggingface.co/KoboldAI/LLaMA2-13B-Tiefighter) |
|
|
|
### Configuration |
|
|
|
The following YAML configuration was used to produce this model: |
|
|
|
```yaml |
|
slices: |
|
- sources: |
|
- model: KoboldAI/LLaMA2-13B-Tiefighter |
|
layer_range: [0, 40] |
|
- model: microsoft/Orca-2-13b |
|
layer_range: [0, 40] |
|
merge_method: slerp |
|
base_model: microsoft/Orca-2-13b |
|
parameters: |
|
t: |
|
- filter: self_attn |
|
value: [0, 0.5, 0.3, 0.7, 1] |
|
- filter: mlp |
|
value: [1, 0.5, 0.7, 0.3, 0] |
|
- value: 0.5 |
|
dtype: bfloat16 |
|
|
|
``` |
