Qiskit/granite-3.2-8b-qiskit
The granite-3.2-8b-qiskit is an 8 billion parameter model developed by IBM Quantum & IBM Research, extended and fine-tuned from granite3.1-8b-base. It specializes in generating high-quality, non-deprecated Qiskit code, trained with Qiskit version 2.0 data. This model is optimized for quantum computing code generation, making it suitable for both experienced practitioners and new Qiskit users.
Loading preview...
Model Overview
The granite-3.2-8b-qiskit is an 8 billion parameter model developed by IBM Quantum & IBM Research. It is an extension and fine-tuning of the granite3.1-8b-base model, specifically designed to enhance capabilities in generating high-quality and non-deprecated Qiskit code. The model was trained exclusively on data with permissive licenses (Apache 2.0, MIT, Unlicense, Mulan PSL v2, BSD-2, BSD-3, CC Attribution 4.0) and is compatible with Qiskit version 2.0 APIs and syntax.
Key Capabilities
- Qiskit Code Generation: Excels at producing functional Qiskit code for quantum computing tasks.
- Instruction Following: Capable of responding to Qiskit coding-related instructions and questions.
- Qiskit v2.0 Compatibility: Ensures generated code adheres to the latest Qiskit standards.
Training and Data
The model's training data includes publicly available code datasets (e.g., GitHub) and synthetic data generated at IBM Quantum, with code older than 2023 excluded. Rigorous data processing involved exact and fuzzy deduplication, as well as filtering for HAP, PII, and malware, leveraging the base model's initial filtering. Training was conducted on IBM's supercomputing cluster (Vela) using NVIDIA A100 GPUs.
Benchmarks and Performance
While not leading in all benchmarks, granite-3.2-8b-qiskit demonstrates competitive performance in specific areas. For instance, it achieves 9.93 on QiskitHumanEval-Hard and 24.50 on QiskitHumanEval, alongside 60.79 on IFEval and 40.51 on TruthfulQA (MC1 acc).
Intended Use Cases
This model is primarily intended as an assistant for:
- Quantum Computing Practitioners: Aiding in the development and generation of Qiskit code.
- New Qiskit Users: Providing support for learning and building Qiskit programs.
Limitations and Ethical Considerations
Users should be aware that generated code is not guaranteed to be flawless and the model has not undergone safety alignment, potentially producing problematic outputs. There is also an acknowledged risk of hallucination, particularly in smaller models, and the community is urged to use the model ethically and responsibly.