Artificial-Development-Studios/QuantumAI-granite-3.3-8b-qiskit

Model Overview

granite-3.3-8b-qiskit is an 8 billion parameter language model developed by IBM Quantum & IBM Research. It is built upon the granite3.3-8b-base model and has been extensively fine-tuned using Qiskit code and instruction data. The primary goal of this specialization is to significantly improve its ability to generate accurate, high-quality, and non-deprecated Qiskit code, ensuring compatibility with Qiskit version 2.0 APIs and syntax.

Key Capabilities

• Quantum Code Generation: Excels at generating quantum computing code using the Qiskit framework.
• Instruction Following: Capable of responding to Qiskit coding-related instructions and questions.
• Qiskit 2.0 Compatibility: Trained specifically with Qiskit version 2.0 data, ensuring generated code adheres to current standards.

Training and Data

The model's training data includes publicly available code from sources like GitHub and synthetic data generated at IBM Quantum. This data undergoes rigorous filtering, including exact and fuzzy deduplication, and PII redaction. The base model's HAP and malware filtering are also leveraged.

Benchmarks

While granite-3.3-8b-qiskit shows strong performance in Qiskit-specific benchmarks, achieving 14.57 on QiskitHumanEval-Hard and 27.15 on QiskitHumanEval, it is noted that other Qiskit-specialized models like qwen2.5-coder-14b-qiskit and mistral-small-3.2-24b-qiskit may outperform it in certain general coding or reasoning benchmarks.

Intended Use

This model is ideal for both experienced quantum computing practitioners and new Qiskit users seeking an assistant for building Qiskit code or answering coding-related queries. Users should be aware that, like all LLMs, it has not undergone safety alignment and may produce problematic outputs, and generated code is not guaranteed to work as intended.