bytesbrains/naderu-loom-7b
The naderu-loom-7b model by Naderu (a BytesBrains Pte. Ltd. venture) is a 7 billion parameter language model fine-tuned from Mistral-7B-Instruct-v0.3. It is specifically designed as a compact, specialized game-master for interactive fiction, generating narrative scenes and structured JSON turns based on game state and player actions. This model excels at producing consistent, schema-compliant JSON output for offline, on-device game narration.
Loading preview...
Overview
naderu-loom-7b is a specialized 7 billion parameter model developed by Naderu, a BytesBrains Pte. Ltd. venture. It is fine-tuned from mistralai/Mistral-7B-Instruct-v0.3 using QLoRA on a 4-bit MLX quant, with weights fused and de-quantized to bf16 for portability. The model's primary function is to act as an offline interactive-fiction narrator, or game master, designed to run efficiently on-device.
Key Capabilities
- Interactive Fiction Narration: Generates narrative scenes and player choices based on game state and actions.
- Structured JSON Output: Produces a single JSON object per turn, including
scenenarration,choices, andstate_deltafor application rendering and state updates. - Schema Compliance: Achieves a high rate of valid JSON output (1.000) and schema compliance (0.971) based on internal evaluations.
- World-State Tracking: Demonstrates ability to track game state, such as inventory, with zero state violations in evaluations.
- Compact and Efficient: Optimized for offline, on-device execution, trained with MLX on an Apple M4 Mac Mini.
Use Cases
This model is ideal for developers creating interactive fiction games or applications that require a compact, specialized AI to generate dynamic narratives and structured game turns. Its focus on consistent JSON output and on-device performance makes it suitable for mobile or embedded gaming environments where offline capabilities are crucial.