This is Mistral, but in 11B, and finetuned with qlora of Pippa and LimaRPv3 dataset.
I took layers of the original Mistral-7B, and duplicated some layer, this is the first frankeinstein method that I found "acceptable" to expend Mistral.
It seems that the first 8 layers of the model is very important, having duplicate of those layers in the model make me think it confuse the model.
UPDATE: Forced mergekit to output bfloat16 file, should be the same thing, but since the base model is bfloat16, wanted it to stay bf16 like the OG model. Even if it was written bfloat16 in the config file earlier, it was float16.
Description
This repo contains quantized files of Mistral-11B-RP-v1.
Model and loras used
- Mistral-7B-v0.1
- PIPPA dataset 11B qlora made on the spot and trained for this model specifically
- LimaRPv3 dataset 11B qlora made on the spot and trained for this model specifically
Prompt template: Alpaca
Below is an instruction that describes a task. Write a response that appropriately completes the request.
### Instruction:
{prompt}
### Response:
The secret sauce
slices:
- sources:
- model: mistralai/Mistral-7B-v0.1
layer_range: [0, 24]
- sources:
- model: mistralai/Mistral-7B-v0.1
layer_range: [8, 32]
merge_method: passthrough
dtype: bfloat16
Special thanks to Sushi.
If you want to support me, you can here.
- Downloads last month
- 5