Update README.md
Browse files
README.md
CHANGED
|
@@ -22,10 +22,38 @@ datasets:
|
|
| 22 |
- Mihaiii/OpenHermes-2.5-1k-longest-curated
|
| 23 |
language:
|
| 24 |
- en
|
|
|
|
| 25 |
library_name: transformers
|
| 26 |
tags:
|
| 27 |
- code
|
| 28 |
---
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 29 |
# Model Card for NinjaMouse-2.4B-32L-danube
|
| 30 |
|
| 31 |
A lanky version of [h2o-danube](https://huggingface.co/h2oai/h2o-danube-1.8b-chat)'s tiny language model, stretched from 24 layers to 32. I have done this in steps, adding 2 new layers per step and training them on different datasets. This seems to have made it a quick learner, and easily fit an 8GB GPU for finetuning using Unsloth for optimizations. This model is designed to be a gateway into bigger language models.
|
|
|
|
| 22 |
- Mihaiii/OpenHermes-2.5-1k-longest-curated
|
| 23 |
language:
|
| 24 |
- en
|
| 25 |
+
inference: false
|
| 26 |
library_name: transformers
|
| 27 |
tags:
|
| 28 |
- code
|
| 29 |
---
|
| 30 |
+
# NinjaMouse-2.4B-32L-danube-exl2
|
| 31 |
+
Original model: [NinjaMouse-2.4B-32L-danube](https://huggingface.co/trollek/NinjaMouse-2.4B-32L-danube)
|
| 32 |
+
Model creator: [trollek](https://huggingface.co/trollek)
|
| 33 |
+
|
| 34 |
+
## Quants
|
| 35 |
+
[4bpw h6](https://huggingface.co/cgus/NinjaMouse-2.4B-32L-danube-exl2/tree/main)
|
| 36 |
+
[4.25bpw h6](https://huggingface.co/cgus/NinjaMouse-2.4B-32L-danube-exl2/tree/4.25bpw-h6)
|
| 37 |
+
[4.65bpw h6](https://huggingface.co/cgus/NinjaMouse-2.4B-32L-danube-exl2/tree/4.65bpw-h6)
|
| 38 |
+
[5bpw h6](https://huggingface.co/cgus/NinjaMouse-2.4B-32L-danube-exl2/tree/5bpw-h6)
|
| 39 |
+
[6bpw h6](https://huggingface.co/cgus/NinjaMouse-2.4B-32L-danube-exl2/tree/6bpw-h6)
|
| 40 |
+
[8bpw h8](https://huggingface.co/cgus/NinjaMouse-2.4B-32L-danube-exl2/tree/8bpw-h8)
|
| 41 |
+
|
| 42 |
+
## Quantization notes
|
| 43 |
+
Made with exllamav2 0.0.15 with default dataset. I'm very unsure about context size of this model.
|
| 44 |
+
For me it breaks past 3000 context, at about 3500 or so. Both with these quants and the creator's GGUF files.
|
| 45 |
+
At first I thought I had some quantization issues but it's probably just the model itself.
|
| 46 |
+
|
| 47 |
+
## How to run
|
| 48 |
+
This quantization method uses GPU and requires Exllamav2 loader which can be found in following applications:
|
| 49 |
+
|
| 50 |
+
[Text Generation Webui](https://github.com/oobabooga/text-generation-webui)
|
| 51 |
+
|
| 52 |
+
[KoboldAI](https://github.com/henk717/KoboldAI)
|
| 53 |
+
|
| 54 |
+
[ExUI](https://github.com/turboderp/exui)
|
| 55 |
+
|
| 56 |
+
# Original model card
|
| 57 |
# Model Card for NinjaMouse-2.4B-32L-danube
|
| 58 |
|
| 59 |
A lanky version of [h2o-danube](https://huggingface.co/h2oai/h2o-danube-1.8b-chat)'s tiny language model, stretched from 24 layers to 32. I have done this in steps, adding 2 new layers per step and training them on different datasets. This seems to have made it a quick learner, and easily fit an 8GB GPU for finetuning using Unsloth for optimizations. This model is designed to be a gateway into bigger language models.
|