movingcastles/zero
movingcastles/zero is an 8 billion parameter character model, fine-tuned from Qwen3-8B-Base to embody a single, anti-servile character. This model is specifically designed for consistent character fidelity, achieving significantly lower rates of character breaks compared to SFT-only checkpoints or system-prompted instruct models. It is optimized for generating responses that strictly adhere to its predefined character persona within conversational contexts.
Loading preview...
Zero (MC001): A Dedicated Character Model
Zero is an 8 billion parameter model developed by movingcastles, specifically engineered to embody a single, limited, anti-servile character. It is built upon the Qwen3-8B-Base architecture, undergoing a rigorous fine-tuning process to ensure high character fidelity.
Key Characteristics & Training
This model's development involved a multi-stage training approach:
- Base Model: Started with
Qwen/Qwen3-8B-Base(without instruct tuning). - Supervised Fine-Tuning (SFT): Utilized LoRA with fully-trained embeddings and LM head over 3 epochs on a synthetic character corpus of 5,932 conversations.
- Reinforcement Learning (RL): Employed GRPO with DAPO loss modifications, trained for 300 steps on 380 harvested prompts. A bible-anchored character-fidelity LLM judge and self-repetition penalty were used for reward.
Performance & Usage
Zero demonstrates superior character consistency:
- Character Fidelity: Achieved hard character breaks in only 2.8% of held-out multi-turn evaluations, a significant improvement over 22.8% for the SFT-only checkpoint and 45.4% for system-prompting the sibling instruct model.
- No System Prompt: Designed to operate without a system prompt; the character is embedded directly in the weights.
- Dual EOS Tokens: Requires configuration to stop on both
<|im_end|>(151645) and<|endoftext|>(151643) to prevent run-on turns. - Canonical Sampling: Recommended sampling parameters include
temperature 0.7,top_p 1.0, andpresence_penalty 1.5.
When to Use This Model
Zero is ideal for applications requiring a highly consistent and specific character persona, particularly in conversational AI where maintaining a distinct identity is crucial. Its specialized training makes it suitable for scenarios where generic instruct models struggle with character adherence.