OpenAI has launched fine-tuning capabilities for its GPT-4o model, a new feature now available to developers using the paid tiers. With this development, the company aims to enhance the AI's adaptability and precision for specific applications.
The fine-tuning function lets developers use custom datasets to optimize GPT-4o for their particular needs. This allows modifications to the response structure and tone, as well as improved adherence to complex and niche-specific directives. OpenAI's adjustments can notably boost the model's efficiency in various sectors, from software development to creative content generation, often requiring just a small dataset for training.
Details on Costs and Accessibility
The fine-tuning feature is immediately accessible, with training priced at $25 for every million tokens. Inference costs are set at $3.75 per million input tokens and $15 per million output tokens. An incentive is available where OpenAI offers 1 million training tokens daily for free to each organization until September 23. Additionally, the GPT-4o mini fine-tuning option grants 2 million free tokens per day during the same period.
Before its public release, OpenAI conducted trials with several partners. For instance, Cosine's Genie, an AI assistant for software engineering, achieved scores of 43.8% on the Verified benchmark and 30.08% on SWE-bench Full. Another early adopter, Distyl, an AI solutions provider, secured the top spot on the BIRD-SQL benchmark, attaining an execution accuracy of 71.83%, particularly excelling in tasks like query reformulation and SQL generation.
Privacy and Security Measures
The fine-tuned models are wholly controlled by the users, who retain ownership of all their data, including inputs and outputs. OpenAI has incorporated multiple layers of safety features to mitigate risks, with ongoing automated safety assessments and monitoring to ensure compliance with its policies.
In July, OpenAI also started introducing an advanced voice feature for its ChatGPT platform, which leverages GPT-4o. This enables highly realistic audio responses, facilitating real-time, smooth conversations with ChatGPT. The alpha version of this Advanced Voice Mode was initially made available to a select group of ChatGPT Plus subscribers, with a broader rollout planned for later.
