metadata
license: other
license_name: deepseek
license_link: https://github.com/deepseek-ai/DeepSeek-Coder/blob/main/LICENSE-MODEL
DeepMagic-Coder-7b
Alternate version:
This is an extremely successful merge of the deepseek-coder-6.7b-instruct and Magicoder-S-DS-6.7B models, bringing an uplift in overall coding performance without any compromise to the models integrity (at least with limited testing).
This is the first of my models to use the merge-kits task_arithmetic merging method. The method is detailed bellow, and its clearly very usefull for merging ai models that were fine-tuned from a common base:
Task Arithmetic:
Computes "task vectors" for each model by subtracting a base model.
Merges the task vectors linearly and adds back the base.
Works great for models that were fine tuned from a common ancestor.
Also a super useful mental framework for several of the more involved
merge methods.
The original models used in this merge can be found here:
The Merge was created using Mergekit and the paremeters can be found bellow:
models:
- model: deepseek-ai_deepseek-coder-6.7b-instruct
parameters:
weight: 1
- model: ise-uiuc_Magicoder-S-DS-6.7B
parameters:
weight: 1
merge_method: task_arithmetic
base_model: ise-uiuc_Magicoder-S-DS-6.7B
parameters:
normalize: true
int8_mask: true
dtype: float16