npow/in-character-rp-12b-v0.1

TEXT GENERATIONConcurrency Cost:1Model Size:12BQuant:FP8Ctx Length:32kTool Calling:SupportedPublished:Jun 25, 2026License:apache-2.0Architecture:Transformer Open Weights Cold

The npow/in-character-rp-12b-v0.1 model is a 12 billion parameter language model, fine-tuned from Mistral-Nemo-Base-2407, specifically designed for roleplay and creative writing. It excels at maintaining consistent character personalities and narrative integrity across conversations, preventing characters from deviating from their established traits. This model is optimized for scenarios requiring distinct and persistent character voices, offering a 32768 token context length for extended interactions.

Loading preview...

npow/in-character-rp-12b-v0.1: In-Character Roleplay Model

This model is a 12 billion parameter fine-tune of Mistral-Nemo-Base-2407, developed by npow, with a primary focus on in-character roleplay and creative writing. Its core strength lies in ensuring characters maintain their distinct personalities and traits throughout a conversation, preventing "positivity drift" or characters opening up prematurely.

Key Capabilities

  • Consistent Character Portrayal: Designed to keep villains villainous and guarded characters reserved, preserving narrative integrity.
  • Creative Writing: Optimized for generating text that adheres to established character personas and story arcs.
  • Extended Context: Supports a 32768 token context length, suitable for longer roleplay sessions and detailed narratives.
  • ChatML Format: Utilizes the ChatML prompt format for structured interactions, including system messages for character cards.

Training Details

The model was trained using QLoRA (4-bit) and high-rank rsLoRA (r=256) on a diverse dataset including LimaRP, PIPPA, CoSER, Gryphe Sonnet-3.5 character cards, WritingPrompts, and Opus-Instruct. The training data was specifically filtered to remove refusals, positivity-drift, and "slop," ensuring a high-quality base for character consistency. This v0.1 release serves as the supervised fine-tuning (SFT) base, with further advancements planned for v0.2 to incorporate diversity-preserving DPO and anti-slop FTPO.

Recommended Usage

For optimal performance, the model recommends specific sampling parameters such as temperature 1.0, min_p 0.05, and DRY (multiplier 0.8, base 1.75, allowed_length 2).

Finetunes
1 models