granite-3.2-8b-qiskit is a 8B parameter model extend pretrained and fine tuned on top of granite3.1-8b-base using Qiskit code and instruction data to improve capabilities at writing high-quality and non-deprecated Qiskit code. We used only data with the following licenses: Apache 2.0, MIT, the Unlicense, Mulan PSL Version 2, BSD-2, BSD-3, and Creative Commons Attribution 4.0.
The model has been trained with Qiskit version 2.0, ensuring compatibility with its APIs and syntax.
This model is designed for generating quantum computing code using Qiskit. Both quantum computing practitionners and new Qiskit users can use this model as an assistant for building Qiskit code or responding to Qiskit coding related instructions and questions.
Generation
This is a simple example of how to use granite-3.2-8b-qiskit model.
python
1import torch
2from transformers import AutoModelForCausalLM, AutoTokenizer
3device ="cuda"# or "cpu"4model_path ="qiskit/granite-3.2-8b-qiskit"5tokenizer = AutoTokenizer.from_pretrained(model_path)6# drop device_map if running on CPU7model = AutoModelForCausalLM.from_pretrained(model_path, device_map=device)8model.eval()9# change input text as desired10chat =[11{"role":"user","content":"Build a random circuit with 5 qubits"},12]13chat = tokenizer.apply_chat_template(chat, tokenize=False, add_generation_prompt=True)14# tokenize the text15input_tokens = tokenizer(input_text, return_tensors="pt")16# move tokenized inputs to device17for i in input_tokens:18 input_tokens[i]= input_tokens[i].to(device)19# generate output tokens20output = model.generate(**input_tokens, max_new_tokens=128)21# decode output tokens into text22output = tokenizer.batch_decode(output)23# loop over the batch to print, in this example the batch size is 124for i in output:25print(i)
Comparison of Qiskit models across benchmarks
Model
QiskitHumanEval-Hard
QiskitHumanEval
HumanEval
ASDiv
MathQA
SciQ
MBPP
IFEval
CrowsPairs (English)
TruthfulQA (MC1 acc)
qwen2.5-coder-14b-qiskit
25.17
49.01
91.46
4.21
53.90
97.00
77.60
49.64
65.18
37.82
mistral-small-3.2-24b-qiskit
32.45
47.02
77.49
3.77
49.68
97.50
64.00
48.44
67.08
39.41
granite-3.3-8b-qiskit
14.57
27.15
62.80
0.48
38.66
93.30
52.40
59.71
59.75
39.05
granite-3.2-8b-qiskit
9.93
24.50
57.32
0.09
41.41
96.30
51.80
60.79
66.79
40.51
Note: All models listed in the benchmark table were evaluated using their respective system prompt, defined in their Hugging Face model.
Training Data
Data Collection and Filtering: Our code data is sourced from a combination of publicly available datasets (e.g., Code available on https://github.com), and additional synthetic data generated at IBM Quantum. We exclude code that is older than 2023.
Exact and Fuzzy Deduplication: We use both exact and fuzzy deduplication to remove documents having (near) identical code content.
HAP, PII, Malware Filtering: We rely on the base model ibm-granite/granite-3.1-8b-code-base for HAP and malware filtering from the initial datasets used in the context of the base model. We also make sure to redact Personally Identifiable Information (PII) in our datasets by replacing PII content (e.g., names, email addresses, keys, passwords) with corresponding tokens (e.g., ⟨NAME⟩, ⟨EMAIL⟩, ⟨KEY⟩, ⟨PASSWORD⟩).
Infrastructure
We train granite-3.2-8b-qiskit using IBM's super computing cluster (Vela) using NVIDIA A100 GPUs. The cluster provides a scalable and efficient infrastructure for training.
Ethical Considerations and Limitations
The use of Large Language Models involves risks and ethical considerations people must be aware of. Regarding code generation, caution is urged against complete reliance on specific code models for crucial decisions or impactful information as the generated code is not guaranteed to work as intended. granite-3.2-8b-qiskit model is not the exception in this regard. Even though this model is suited for multiple code-related tasks, it has not undergone any safety alignment, there it may produce problematic outputs. Additionally, it remains uncertain whether smaller models might exhibit increased susceptibility to hallucination in generation scenarios by copying source code verbatim from the training dataset due to their reduced sizes and memorization capacities. This aspect is currently an active area of research, and we anticipate more rigorous exploration, comprehension, and mitigations in this domain. Regarding ethics, a latent risk associated with all Large Language Models is their malicious utilization. We urge the community to use granite-3.2-8b-qiskit model with ethical intentions and in a responsible way.
