|
--- |
|
base_model: [] |
|
library_name: transformers |
|
tags: |
|
- mergekit |
|
- merge |
|
- llama |
|
- not-for-all-audiences |
|
--- |
|
|
|
# GGUF / IQ / Imatrix for [Silver-Sun-11B](https://huggingface.co/ABX-AI/Silver-Sun-11B) |
|
|
|
 |
|
|
|
**RE-UPLOAD: The configuration was wrong on the previous quantization. Fixed now! All quants are re-uploaded and Q8 is added** |
|
|
|
**Why Importance Matrix?** |
|
|
|
**Importance Matrix**, at least based on my testing, has shown to improve the output and performance of "IQ"-type quantizations, where the compression becomes quite heavy. |
|
The **Imatrix** performs a calibration, using a provided dataset. Testing has shown that semi-randomized data can help perserve more important segments as the compression is applied. |
|
|
|
Related discussions in Github: |
|
[[1]](https://github.com/ggerganov/llama.cpp/discussions/5006) [[2]](https://github.com/ggerganov/llama.cpp/discussions/5263#discussioncomment-8395384) |
|
|
|
The imatrix.txt file that I used contains general, semi-random data, with some custom kink. |
|
|
|
# Silver-Sun-11B |
|
|
|
> I'd like to experiment more with merging 11B, hopefully adding more options of this weight class. |
|
> This model is good at writing and reasoning, with a preference for more profane NSFW language when the appropriate cards are used. |
|
> I've been having fun with it so far, although at times it can be a bit blunt, although some may prefer that. It's also highly uncensored. |
|
|
|
Works best with Alpaca instruction presets. |
|
|
|
## Merge Details |
|
|
|
This is a merge of pre-trained language models created using [mergekit](https://github.com/cg123/mergekit). |
|
|
|
### Merge Method |
|
|
|
This model was merged using the SLERP merge method. |
|
|
|
### Models Merged |
|
|
|
The following models were included in the merge: |
|
* ABX-AI/Solstice-FKL-11B |
|
>[!NOTE] |
|
>A mixture of [Sao10K/Solstice-11B-v1](https://huggingface.co/Sao10K/Solstice-11B-v1) and [saishf/Fimbulvetr-Kuro-Lotus-10.7B](https://huggingface.co/saishf/Fimbulvetr-Kuro-Lotus-10.7B) |
|
* [Himitsui/Kaiju-11B](https://huggingface.co/Himitsui/Kaiju-11B) |
|
|
|
### OpenLLM Eval Results |
|
|
|
[Detailed Results + Failed GSM8K](https://huggingface.co/datasets/open-llm-leaderboard/details_ABX-AI__Silver-Sun-11B) |
|
|
|
|
|
>[!NOTE] |
|
>I had to remove GSM8K from the results and manually re-average the rest. GSM8K failed due to some issue with formatting, which is not experienced during practical usage. |
|
>By removing the GSM8K score, the average is VERY close to upstage/SOLAR-10.7B-v1.0 (74.20), which would make sense. |
|
>Feel free to ignore the actual average and use the other scores individually for reference. |
|
|
|
| Metric |Value| |
|
|---------------------------------|----:| |
|
|Avg. |74.13| |
|
|AI2 Reasoning Challenge (25-Shot)|69.80| |
|
|HellaSwag (10-Shot) |87.91| |
|
|MMLU (5-Shot) |66.90| |
|
|TruthfulQA (0-shot) |61.89| |
|
|Winogrande (5-shot) |84.14| |
|
|
|
### Configuration |
|
|
|
The following YAML configuration was used to produce this model: |
|
|
|
```yaml |
|
slices: |
|
- sources: |
|
- model: ABX-AI/Solstice-FKL-11B |
|
layer_range: [0, 48] |
|
- model: Himitsui/Kaiju-11B |
|
layer_range: [0, 48] |
|
merge_method: slerp |
|
base_model: ABX-AI/Solstice-FKL-11B |
|
parameters: |
|
t: |
|
- filter: self_attn |
|
value: [0.6, 0.7, 0.8, 0.9, 1] |
|
- filter: mlp |
|
value: [0.4, 0.3, 0.2, 0.1, 0] |
|
- value: 0.5 |
|
dtype: bfloat16 |
|
|
|
``` |
