68 lines
3.9 KiB
Markdown
68 lines
3.9 KiB
Markdown
---
|
|
base_model:
|
|
- inflatebot/MN-12B-Mag-Mell-R1
|
|
- TheDrummer/UnslopNemo-12B-v4.1
|
|
library_name: transformers
|
|
tags:
|
|
- mergekit
|
|
- merge
|
|
- 12b
|
|
- chat
|
|
- roleplay
|
|
- creative-writing
|
|
- SLERP
|
|
license: apache-2.0
|
|
new_version: redrix/patricide-12B-Unslop-Mell-v2
|
|
---
|
|
# patricide-12B-Unslop-Mell
|
|
>The sins of the Father shan't ever be repeated this way.
|
|
|
|
|
|
|
|
This is a merge of pre-trained language models created using [mergekit](https://github.com/cg123/mergekit).
|
|
|
|
This is my first merge — I still have no idea how writing the parameters in the config actually works. (**Update:** I figured it out) If anyone has more extensive guides for merging, please let me know. I would also like to get into the science behind all this.
|
|
|
|
Both models produced enjoyable results, so I decided to merge them, to create a model hopefully inheriting good traits of the parents. (**Update:** The early testing of this model revealed good coherency, but it sometimes spits out unintelligeble gibberish or made-up words. This is likely due to the broken tokenizer)
|
|
|
|
I've tested this model on the [**Q6_K GGUF**](https://huggingface.co/redrix/patricide-12B-Unslop-Mell-GGUF/blob/main/patricide-12B-Unslop-Mell-Q6_K.gguf) Quant and it provided satisfactory results, thus I decided to upload it. Although I've not extensively tested it in Storywriting nor RP, the results were stable and *at least* coherent. I tested it on a **Temperature of 1** (Temperature last) and **Min-P of 0.1**. I don't know the effects **DRY** or **XTC** have on the stability of the output, or how it fares on high context sizes. Both parent models use the **ChatML** Template. Although [Unslop-Nemo](https://huggingface.co/TheDrummer/UnslopNemo-12B-v4.1) also uses **Metharme/Pygmalion**. I've not yet tested which works better. (**Update:** Mergekit introduced a feature to define the template; I will force it to use ChatML in my next models, so it has an all-around standard.)
|
|
|
|
Feel free to experiment, as I am only experimenting myself.
|
|
|
|
**Update:** I will likely release my next models once I am able to run them, without too much fine-tuning of samplers/parameters/text templates/etc. Extensive testing as per [DavidAU's approach](https://huggingface.co/DavidAU/Maximizing-Model-Performance-All-Quants-Types-And-Full-Precision-by-Samplers_Parameters) will be done afterwards, so I may gain more impressions, while being able to work on new models already. I would like to create models that are very good in their base states, with samplers being the thing to perfect them. As such I won't spend too much time finetuning samplers, unless the model's base state is very promising.
|
|
|
|
# Quantization
|
|
Static **GGUF** Quants available at:
|
|
- [redrix/patricide-12B-Unslop-Mell-GGUF](https://huggingface.co/redrix/patricide-12B-Unslop-Mell-GGUF) (has less quants than below ⬇️)
|
|
- [mradermacher/patricide-12B-Unslop-Mell-GGUF](https://huggingface.co/mradermacher/patricide-12B-Unslop-Mell-GGUF) (Thanks ♥️)
|
|
|
|
Weighted/Imatrix **GGUF** Quants available at [mradermacher/patricide-12B-Unslop-Mell-i1-GGUF](https://huggingface.co/mradermacher/patricide-12B-Unslop-Mell-i1-GGUF).
|
|
|
|
## Merge Details
|
|
### Merge Method
|
|
|
|
This model was merged using the SLERP merge method.
|
|
|
|
### Models Merged
|
|
|
|
The following models were included in the merge:
|
|
* [inflatebot/MN-12B-Mag-Mell-R1](https://huggingface.co/inflatebot/MN-12B-Mag-Mell-R1)
|
|
* [TheDrummer/UnslopNemo-12B-v4.1](https://huggingface.co/TheDrummer/UnslopNemo-12B-v4.1)
|
|
|
|
### Configuration
|
|
|
|
The following YAML configuration was used to produce this model:
|
|
|
|
```yaml
|
|
models:
|
|
- model: TheDrummer/UnslopNemo-12B-v4.1
|
|
- model: inflatebot/MN-12B-Mag-Mell-R1
|
|
merge_method: slerp
|
|
base_model: TheDrummer/UnslopNemo-12B-v4.1
|
|
dtype: bfloat16
|
|
parameters:
|
|
t: [0, 0.5, 1, 0.5, 0]
|
|
|
|
```
|
|
|
|
> I made the cover art myself in Photoshop... I don't use AI for stuff like that. |