Downloads · 30 days
6
12% of all-time downloads
RichardErkhov/Blizado_-_discolm-mfto-7b-german-v0.1-8bits
Blizado_-_discolm-mfto-7b-german-v0.1-8bits is a machine learning model from RichardErkhov. Use it for the machine learning task on the model card, and read the license before you ship it in a product.
Downloads · 30 days
6
12% of all-time downloads
All-time downloads
51
Public
Parameters
7.2B
7.5 GB on disk
Likes
0
Public
Click a slice to open those files.
.safetensors7.5 GB · 100%
How the weights are stored.
I87B · 96%
From the Hugging Face model README
Quantization made by Richard Erkhov.
discolm-mfto-7b-german-v0.1 - bnb 8bits
base_model: [] tags:
This is a experimental merge of pre-trained language models created using mergekit.
This model was merged using the SLERP merge method.
The following models were included in the merge:
Because both used models are, to my knowledge, the two best models when it comes to German language generation.
DiscoLM German 7B is is up to this date (01/21/2024) by far the best German model and makes far fewer grammatical errors and his German generally sounds good. But it is finetuned on Mistral V0.2 or even V0.1.
Mistral FT Optimized 1227 is much better in German than Mistral 7B V0.2 and other German fine-tuning models that make grammar errors in almost every sentence. But even that model is a good step behind DiscoLM German 7B and creates not so well formed German sentences.
The ulterior motive was now combining this two models to get a even better German model, especially for German roleplay.
The following YAML configuration was used to produce this model:
slices:
- sources:
- model: OpenPipe/mistral-ft-optimized-1227
layer_range: [0, 32]
- model: DiscoResearch/DiscoLM_German_7b_v1
layer_range: [0, 32]
merge_method: slerp
base_model: OpenPipe/mistral-ft-optimized-1227
parameters:
t:
- value: [0.5, 0.9]
dtype: bfloat16
This settings are from the model oshizo/japanese-e5-mistral-7b_slerp.