91 lines
2.5 KiB
Markdown
91 lines
2.5 KiB
Markdown
|
---
|
||
|
|
base_model:
|
||
|
|
- allura-org/TQ2.5-14B-Neon-v1
|
||
|
|
- allura-org/TQ2.5-14B-Sugarquill-v1
|
||
|
|
library_name: transformers
|
||
|
|
tags:
|
||
|
|
- mergekit
|
||
|
|
- merge
|
||
|
|
license: apache-2.0
|
||
|
|
language:
|
||
|
|
- en
|
||
|
|
---
|
||
|
|
|
||
|
|
<img src="aletheia.png">
|
||
|
|
<small>Image by CalamitousFelicitousness</small>
|
||
|
|
|
||
|
|
---
|
||
|
|
|
||
|
|
|
||
|
|
# Qwen2.5-14B Aletheia v1
|
||
|
|
|
||
|
|
RP/Story hybrid model, merge of Sugarquill and Neon. As with Gemma version, I wanted to preserve Sugarquill's creative spark, while making the model more steerable for RP. It proved to be more difficult this time, but I quite like the result regardless, even if the model is still somewhat temperamental.
|
||
|
|
|
||
|
|
Should work for both RP and storywriting, either on raw completion or with back-and-forth cowriting in chat mode. Seems to be quite sensitive to low depth instructions and samplers.
|
||
|
|
|
||
|
|
Thanks to Toasty and Fizz for testing and giving feedback
|
||
|
|
|
||
|
|
Model was created by Auri.
|
||
|
|
|
||
|
|
---
|
||
|
|
|
||
|
|
**Notes about merging**
|
||
|
|
|
||
|
|
It took me 20 something attempts to make this model. TIES didn't work at all, producing broken or nearly broken results every time. SLERP worked much better and after just 3 attempts I got something I like.
|
||
|
|
Sugarquill was really prone to overtaking the merge, so I had to reduce it's part a lot, and still model has a lot of influence from it.
|
||
|
|
|
||
|
|
**Format**
|
||
|
|
|
||
|
|
Model responds to ChatML instruct formatting, exactly like it's base model.
|
||
|
|
|
||
|
|
```
|
||
|
|
<|im_start|>system
|
||
|
|
{system message}<|im_end|>
|
||
|
|
<|im_start|>user
|
||
|
|
{user message}<|im_end|>
|
||
|
|
<|im_start|>assistant
|
||
|
|
{response}<|im_end|>
|
||
|
|
```
|
||
|
|
|
||
|
|
**Recommended Samplers**
|
||
|
|
|
||
|
|
This one is a bit of a special snowflake, with special tastes. Those seem to work pretty well:
|
||
|
|
|
||
|
|
```
|
||
|
|
Temperature - 0.8
|
||
|
|
Top-A - 0.3
|
||
|
|
TFS - 0.75
|
||
|
|
DRY - Multiplier 0.8 - Base 1.75 - Allowed length 3 - Range 1024
|
||
|
|
```
|
||
|
|
|
||
|
|
As a starting point, you can try this [ST Master Import](https://huggingface.co/allura-org/TQ2.5-14B-Aletheia-v1/blob/main/TQ-Aletheia.json)
|
||
|
|
|
||
|
|
## Merge Details
|
||
|
|
### Merge Method
|
||
|
|
|
||
|
|
This model was merged using the SLERP merge method.
|
||
|
|
|
||
|
|
### Models Merged
|
||
|
|
|
||
|
|
The following models were included in the merge:
|
||
|
|
* [allura-org/TQ2.5-14B-Neon-v1](https://huggingface.co/allura-org/TQ2.5-14B-Neon-v1)
|
||
|
|
* [allura-org/TQ2.5-14B-Sugarquill-v1](https://huggingface.co/allura-org/TQ2.5-14B-Sugarquill-v1)
|
||
|
|
|
||
|
|
### Configuration
|
||
|
|
|
||
|
|
The following YAML configuration was used to produce this model:
|
||
|
|
|
||
|
|
```yaml
|
||
|
|
base_model: allura-org/TQ2.5-14B-Sugarquill-v1
|
||
|
|
dtype: bfloat16
|
||
|
|
merge_method: slerp
|
||
|
|
parameters:
|
||
|
|
t:
|
||
|
|
- value: 0.7
|
||
|
|
slices:
|
||
|
|
- sources:
|
||
|
|
- layer_range: [0, 48]
|
||
|
|
model: allura-org/TQ2.5-14B-Neon-v1
|
||
|
|
- layer_range: [0, 48]
|
||
|
|
model: allura-org/TQ2.5-14B-Sugarquill-v1
|
||
|
|
```
|