EphAsad/Mnemosyne-3B

Overview

Mnemosyne-3B is a 3 billion parameter QLoRA fine-tune of the Qwen2.5-Coder-3B-Instruct model, developed by Zain Asad. Its primary function is to translate natural language questions into SQL queries, with a strong specialization in schemas related to laboratory, scientific, food safety, water quality, and environmental microbiology databases. The model is designed for efficient, low-latency local or server-side inference and is available in bf16 precision and GGUF formats (Q4_K_M, Q8_0) for use with tools like llama.cpp, Ollama, and LM Studio.

Key Capabilities

• Domain-Specific Text-to-SQL: Excels at generating SQL for complex laboratory and scientific database schemas, showing a +48% execution accuracy (EX) improvement on a custom lab domain benchmark compared to its base model.
• General-Purpose SQL Generation: Can handle general text-to-SQL tasks, though with a modest regression on cross-domain benchmarks like Spider (-7.8% EX).
• Low-Latency Inference: Optimized for local deployment and quick response times.
• Schema-Aware: Requires a database schema (DDL) to be provided at inference time for accurate query generation.

Intended Use Cases

• Laboratory Information Management Systems (LIMS): Ideal for generating queries within LIMS, food and water testing, and scientific data management applications.
• Developer Tooling & Data Analyst Assistants: Useful for creating tools that assist developers and data analysts in writing SQL queries.
• Schema-Aware Chatbots: Can power chatbots that interact with databases by converting natural language into SQL.

Limitations

• Performance on general SQL tasks is slightly lower than the base model.
• Requires an explicit schema in the prompt; has no inherent knowledge of databases.
• Limited by a 2048-token context length, which may truncate very long schemas.
• Not suited for safety-critical automated execution without human review due to potential for semantically incorrect queries.