metadata
pipeline_tag: text-generation
license: other
license_name: modified-mit
license_link: https://github.com/MiniMax-AI/MiniMax-M2.1/blob/main/LICENSE
base_model:
- MiniMaxAI/MiniMax-M2.1
tags:
- smoothie-qwen
Smoothie-MiniMax-M2.1
Overview
This is a modified version of MiniMax-M2.1, using Smoothie-Qwen.
What is it?
Reduced probability of Kanji, Hanja, Chinese character(radical) tokens to reduce sudden language mixing.
For who?
If you see Chinese characters during non-Chinese conversation, this model will help in this case.
It does not "solve" the main problem, just improve its occurrence.
For Chinese and Japanese users: Use original model! This model will behave worse in these languages.
Result
From my testing:
- Chinese character did not appear on Korean conversation.
- When I ask about Japanese topic, model sucessfully answered with Kanji and Hiragana (although I can't test correctness of response)
How I did it?
I tried to replicate Unsloth's UD quant as possible because my system only can handle up to 3-bit quants.
- Download original model
- Apply Smoothie-qwen (See configs/config.yaml for reference)
- Convert to GGUF (BF16)
- Run llama-quantize with Unsloth imatrix and manual override to tensor type from UD quants
- Run llama-gguf-split (max size 50GB)
Recommendation
At temperature 1.0, tool calling is bit unstable. I recommend temperature=0.7.