Gramscii-IT/SemanticRepair-270M
Gramscii-IT/SemanticRepair-270M is a 270 million parameter model developed by Gramscii-IT, specifically designed to propose request lines for an embedding router or identify 'NO_REQUEST' when no valid request is found. This model does not answer user questions but rather processes natural language inputs to extract and rephrase user intentions into structured request lines for downstream systems. It is fine-tuned from google/gemma-3-270m and excels at accurately identifying and rephrasing user requests, including complex multi-intent queries and handling negations, while maintaining a compact size and efficient CPU-only operation.
Loading preview...
SemanticRepair-270M: Specialized Request Rewriter
Gramscii-IT/SemanticRepair-270M is a compact 270 million parameter model, fine-tuned from google/gemma-3-270m, designed for a very specific task: processing natural language inputs to generate structured request lines for an embedding router, or outputting NO_REQUEST if no valid request is detected. Unlike general-purpose LLMs, this model does not generate answers or execute tools; its sole function is to interpret user intent for routing purposes.
Key Capabilities
- Intent Extraction: Accurately identifies and rephrases user requests into one or more plain request lines.
- Negation Handling: Effectively processes requests containing negations (e.g., "don't search for X, tell me Y").
- Multi-Intent Processing: Can decompose a single message into multiple distinct request lines (e.g., "book a table and cancel my dentist").
- Non-Request Identification: Reliably outputs
NO_REQUESTfor conversational filler, greetings, or non-actionable statements. - CPU-Efficient: Designed for CPU-only operation, demonstrating low latency (e.g., 17-38 ms) and minimal memory footprint (762 MB resident) on a Mac M-series machine.
- Language Support: Trained on five languages (English, Spanish, German, French, Italian) for robust multilingual request parsing.
Good For
- Embedding Routers: Ideal for systems that need to translate diverse user queries into standardized, actionable requests for tool execution or capability admission.
- Pre-processing User Input: Useful as a front-end component to normalize and clarify user intent before it reaches more complex AI agents or backend services.
- Resource-Constrained Environments: Its small size and CPU-only efficiency make it suitable for deployment where GPU resources are limited or latency is critical.
- Reproducible Routing: Employs greedy completion and specific prompt formatting to ensure consistent and deterministic request rephrasing, crucial for reliable routing logic.