初始化项目,由ModelHub XC社区提供模型
Model: TeeZee/Bielik-SOLAR-LIKE-10.7B-Instruct-v0.1 Source: Original Platform
This commit is contained in:
20
README.md
Normal file
20
README.md
Normal file
@@ -0,0 +1,20 @@
|
||||
---
|
||||
license: cc-by-nc-4.0
|
||||
---
|
||||
### TeeZee/Bielik-SOLAR-LIKE-10.7B-Instruct-v0.1 ###
|
||||
Precise recipe used by Upstage to create [SOLAR](https://huggingface.co/upstage/SOLAR-10.7B-v1.0) was applied to https://huggingface.co/speakleash/Bielik-7B-Instruct-v0.1
|
||||
*(just merge, no finetuning)
|
||||
|
||||
### Results ###
|
||||
- model is still coherent in Polish language, even without finetuning after merge
|
||||
- instruct mode works in ooba without issues
|
||||
- model is censored and aligned
|
||||
- seems that this model scores highest amongst all versions of original Bielik models, further finetunig should improve results even more.
|
||||
|
||||

|
||||
|
||||
- on dedicated to Polish speaking LLM leaderboards, its 2nd, just behind instruct version used for this merge, and thats to be expected when applying DUS merge - very small quality loss.
|
||||
|
||||
[Polish LLMs leaderboards](https://huggingface.co/spaces/speakleash/open_pl_llm_leaderboard)
|
||||
|
||||
- overall it seems like a good base for further finetunig in Polish language.
|
||||
Reference in New Issue
Block a user