InferenceIllusionist's picture
Update README.md
43bd337 verified
|
raw
history blame
3.24 kB
---
license: cc-by-nc-4.0
tags:
- GGUF
- iMat
- llama3
---
```
e88 88e d8
d888 888b 8888 8888 ,"Y88b 888 8e d88
C8888 8888D 8888 8888 "8" 888 888 88b d88888
Y888 888P Y888 888P ,ee 888 888 888 888
"88 88" "88 88" "88 888 888 888 888
b
8b,
e88'Y88 d8 888
d888 'Y ,"Y88b 888,8, d88 ,e e, 888
C8888 "8" 888 888 " d88888 d88 88b 888
Y888 ,d ,ee 888 888 888 888 , 888
"88,d88 "88 888 888 888 "YeeP" 888
PROUDLY PRESENTS
```
## Llama-3-8B-EGO-iMat-GGUF
Quantized from fp32 with love.
* Weighted quantizations were calculated using groups_merged.txt with 105 chunks (recommended amount for this file) and n_ctx=512. Special thanks to jukofyork for sharing [this process](https://huggingface.co/jukofyork/WizardLM-2-8x22B-imatrix)
<b>**Note - Please use SillyTavern as well as the following prompt format:**</b>
```
[EGO]Name: Character name and then Everything that forms the personality and speech patterns.(i.e. scenario, sample dialogue, character definitions, etc)[/EGO]
[SEEN]User message.[/SEEN]
Character Name:
```
For a brief rundown of iMatrix quant performance please see this [PR](https://github.com/ggerganov/llama.cpp/pull/5747)
<b>All quants are verified working prior to uploading to repo for your safety and convenience. </b>
Please note importance matrix quantizations are a work in progress. IQ4 and above is recommended for best results.
Original model card [here](https://huggingface.co/Envoid/Llama-3-8B-EGO) and below
---
# This model isn't particularly great. It's just an undercooked experiment.
Releasing it anyways just in case it accidentally makes good merge meat.
# It also has a tendency to produce mature content without warning.
This model is tuned off of the base Llama-3-8B model.
I adapted the leaked Undi dataset into training samples for custom formatting. This model pretty much only functions properly in SillyTavern.
The formatting has two pairs of pseudotokens
```
[EGO]Name: Character name and then Everything that forms the personality and speech patterns.(i.e. scenario, sample dialogue, character definitions, etc)[/EGO]
[SEEN]User message.[/SEEN]
Character Name:
```
The self attention modules were fine tuned separately on this dataset and the pseudotokens were chosen because they made logical sense with respect to the character giving a reply without allowing the model to 'connect the dots' during training and figure out that it is indeed an AI language model.
After this was done all modules were then finetuned together on the dendrite dataset in order to connect the changes made to the attention modules.
So with regards to building a SillyTavern prompt template you basically want the entire story string and any additional stylistic instructions enclosed in the [EGO] tags and then the user messages enclosed in [SEEN] tags.
It doesn't give particularly verbose replies unless you're continueing a roleplay with verbose messages. Otherwise it's pretty bad.