metadata
base_model:
- mistralai/Mistral-Small-3.2-24B-Instruct-2506
datasets:
- ConicCat/AntiRep
Mistral-Small-3.2-AntiRep-24B:
- Exactly what it says on the tin, Orpo'd Mistral Small 3.2 to remove repetition.
- Trained to reduce infinite repetition, repetition of structure and sentences in multi turn conversation, and repetition within responses.
- Got really annoyed with all of my Mistral Small test models having repetition issues, so I decided to whip this up.
- Produced by doing orpo with Qwen 3 8B at 0 temp + .7 rep pen (<1 increases repetition) as rejected vs V3 03/24 as chosen.
- The LoRA is also available too, if you want to use it to reduce repetition on other MS3.2 tunes.
Enjoy!