Text Generation
Transformers
Safetensors
llama
Not-For-All-Audiences
nsfw
text-generation-inference
Inference Endpoints
Description
This repo contains fp16 files of UtopiaXL-13B, a merge I have done with the new layer shuffle method from mergekit (thank you Charles for adding my request to your project!)
This is more a proof of concept showing the following:
- Llama2 is very flexible
- Llama2 don't care about what is finetuned on the layers specifically if you keep them in the same order
- Clean merge (no ties, no SLERP, etc...) with only layer is possible without breaking something
- Deleting special tokens/using model with special token don't break the model
- Alpaca win, always. So use it.
The name "XL" come from the absurd amount of model pushed into it.
Models and loras used
- Undi95/Utopia-13B
- KoboldAI/LLAMA2-13B-Holodeck-1
- Undi95/PsyMedRP-v1-13B
- PygmalionAI/pygmalion-2-13b
- Heralax/Cat-0.5
- KoboldAI/LLaMA2-13B-TiefighterLR
- Heralax/Augmental-13b-two-epochs
- Undi95/Storytelling-v2.1-13B-lora
- Undi95/LimaRP-UtopiaXL-13B-v3-lora
The sauce
!mergekit-layershuffle ./UtopiaXL \
--model Undi95/Utopia-13B --weight 0.4 \
--model KoboldAI/LLAMA2-13B-Holodeck-1 --weight 0.1 \
--model Undi95/PsyMedRP-v1-13B --weight 0.1 \
--model PygmalionAI/pygmalion-2-13b --weight 0.25 \
--model Heralax/Cat-0.5 --weight 0.1 \
--model KoboldAI/LLaMA2-13B-TiefighterLR --weight 0.1 \
--model Heralax/Augmental-13b-two-epochs --weight 0.1 \
--write-yaml UtopiaXL.yaml
=========================
merge_method: passthrough
slices:
- sources:
- layer_range:
- 0
- 1
model: KoboldAI/LLAMA2-13B-Holodeck-1
- sources:
- layer_range:
- 1
- 4
model: Undi95/Utopia-13B
- sources:
- layer_range:
- 4
- 5
model: PygmalionAI/pygmalion-2-13b
- sources:
- layer_range:
- 5
- 6
model: Undi95/PsyMedRP-v1-13B
- sources:
- layer_range:
- 6
- 7
model: KoboldAI/LLaMA2-13B-TiefighterLR
- sources:
- layer_range:
- 7
- 8
model: PygmalionAI/pygmalion-2-13b
- sources:
- layer_range:
- 8
- 9
model: Undi95/PsyMedRP-v1-13B
- sources:
- layer_range:
- 9
- 10
model: PygmalionAI/pygmalion-2-13b
- sources:
- layer_range:
- 10
- 13
model: Undi95/Utopia-13B
- sources:
- layer_range:
- 13
- 14
model: Heralax/Cat-0.5
- sources:
- layer_range:
- 14
- 17
model: Undi95/Utopia-13B
- sources:
- layer_range:
- 17
- 18
model: Heralax/Augmental-13b-two-epochs
- sources:
- layer_range:
- 18
- 19
model: KoboldAI/LLaMA2-13B-TiefighterLR
- sources:
- layer_range:
- 19
- 22
model: PygmalionAI/pygmalion-2-13b
- sources:
- layer_range:
- 22
- 23
model: Undi95/Utopia-13B
- sources:
- layer_range:
- 23
- 25
model: PygmalionAI/pygmalion-2-13b
- sources:
- layer_range:
- 25
- 27
model: Undi95/Utopia-13B
- sources:
- layer_range:
- 27
- 28
model: Heralax/Cat-0.5
- sources:
- layer_range:
- 28
- 30
model: Undi95/Utopia-13B
- sources:
- layer_range:
- 30
- 31
model: Heralax/Augmental-13b-two-epochs
- sources:
- layer_range:
- 31
- 32
model: Undi95/Utopia-13B
- sources:
- layer_range:
- 32
- 33
model: Heralax/Cat-0.5
- sources:
- layer_range:
- 33
- 34
model: Heralax/Augmental-13b-two-epochs
- sources:
- layer_range:
- 34
- 35
model: Undi95/PsyMedRP-v1-13B
- sources:
- layer_range:
- 35
- 36
model: Heralax/Augmental-13b-two-epochs
- sources:
- layer_range:
- 36
- 37
model: Undi95/Utopia-13B
- sources:
- layer_range:
- 37
- 38
model: KoboldAI/LLAMA2-13B-Holodeck-1
- sources:
- layer_range:
- 38
- 39
model: PygmalionAI/pygmalion-2-13b
- sources:
- layer_range:
- 39
- 40
model: KoboldAI/LLAMA2-13B-Holodeck-1
=========================
=> Applying Undi95/Storytelling-v2.1-13B-lora x 0.1
=> Trained on LimaRP for +2h
=> Applying Undi95/LimaRP-UtopiaXL-13B-v3-lora x 0.35
Prompt template: Alpaca
Below is an instruction that describes a task. Write a response that appropriately completes the request.
### Instruction:
{prompt}
### Response:
A big thanks to Charles for adding the layer shuffle method to his tool mergekit and Henky/KoboldAI for the machine he let me use.
If you want to support me, you can here.
- Downloads last month
- 787
This model does not have enough activity to be deployed to Inference API (serverless) yet. Increase its social
visibility and check back later, or deploy to Inference Endpoints (dedicated)
instead.