electroglyph/Qwen3-4B-Instruct-2507-uncensored-unslop
1
0
Fork 0
Code Issues Pull Requests Actions Projects Releases Wiki Activity

初始化项目，由ModelHub XC社区提供模型

Browse Source 
Model: electroglyph/Qwen3-4B-Instruct-2507-uncensored-unslop
Source: Original Platform
This commit is contained in:
ModelHub XC
2026-06-15 21:26:23 +08:00
commit cbde96154a
21 changed files with 152360 additions and 0 deletions
Download Patch File Download Diff File Expand all files Collapse all files

48
README.md Normal file
View File

@@ -0,0 +1,48 @@
---
library_name: transformers
license: apache-2.0
license_link: https://huggingface.co/Qwen/Qwen3-4B-Instruct-2507/blob/main/LICENSE
pipeline_tag: text-generation
---
Update: newer version [here](https://huggingface.co/electroglyph/Qwen3-4B-Instruct-2507-uncensored-unslop-v2)
This is a GSPO finetune to remove slop from this model: [Qwen3-4B-Instruct-2507-uncensored](https://huggingface.co/electroglyph/Qwen3-4B-Instruct-2507-uncensored)
I used the same method (mostly) as this model: [gemma-3-4b-it-unslop-GSPO](https://huggingface.co/electroglyph/gemma-3-4b-it-unslop-GSPO)
Note: This is *not* an RP tune, it's a compliant model with a different style from regular Qwen3 4B 2507.
My uncensoring dataset was generated by Gemma 3 27B abliterated model, which added a lot of Gemma writing style to this model.
It also added some Gemma style slop, which this finetune has *mostly* mitigated...it's probably about 90% of the way there.
There will still be prompts that get quite a bit of the stereotypical LLM slop outputted.
However, I concluded this finetune a bit early because I don't want to damage it too much. I haven't decided yet if this is the final version, but it might be.
I've uploaded a UD-Q4_K_XL GGUF with settings that I grabbed from Unsloth's quant using my lil utility: [quant_clone](https://github.com/electroglyph/quant_clone)
Here are some pics of before and after output with the slop highlighted and total at the bottom:
Prompt = "write a short story about a gothic romance, it should be around 500 words long"
For this model to generate comparable length I had to prompt it for 700 words.
Before:
![before](./orig1.png)
![before](./orig2.png)
![before](./orig3.png)
After:
![before](./new1.png)
![before](./new2.png)
![before](./new3.png)
One thing I noticed: this model generates a bit less text for the same prompt versus the original. I didn't enforce word count for this training run, I'll re-enable it and upload an updated version in a day or so.