GPT-4.5 was trained on a diverse set of datasets, including a mix of publicly available data, proprietary data from data partnerships, and custom datasets developed in-house. These datasets collectively contribute to the model's robust conversational capabilities and world knowledge. However, specific details about the exact datasets used are not explicitly mentioned in the available information.
The training process involved new supervision techniques combined with traditional methods like supervised fine-tuning (SFT) and reinforcement learning from human feedback (RLHF), similar to those used for GPT-4o[1][3]. The model's development also included scalable alignment techniques, where smaller models generate high-quality training data for larger models, enhancing the model's steerability and understanding of nuance[7].
OpenAI's data processing pipeline includes rigorous filtering to maintain data quality and mitigate potential risks. They employ advanced data filtering processes to reduce the processing of personal information when training their models. Additionally, they use a combination of their Moderation API and safety classifiers to prevent the use of harmful or sensitive content[1].
While the exact datasets are not detailed, the approach emphasizes a broad and diverse data foundation to support GPT-4.5's capabilities in tasks such as writing, programming, and solving practical problems with fewer hallucinations[1][3].
Citations:
[1] https://cdn.openai.com/gpt-4-5-system-card.pdf
[2] https://towardsdatascience.com/what-gpt-4-brings-to-the-ai-table-74e392a32ac3/
[3] https://www.lesswrong.com/posts/fqAJGqcPmgEHKoEE6/openai-releases-gpt-4-5
[4] https://www.chatbase.co/blog/gpt-5
[5] https://www.wired.com/story/openai-gpt-45/
[6] https://www.datacamp.com/blog/everything-we-know-about-gpt-5
[7] https://www.vellum.ai/blog/gpt-4-5-is-here-heres-how-good-this-model-is
[8] https://arxiv.org/html/2404.07840v1