OpenAI Data Scientist Interview Questions

OpenAI's models need to access a multitude of features in real-time. Designing a robust feature aggregator supports this need.
Problem statement: Create a class StreamingFeatureAggregator that collects real-time data features for multiple models.
- add_feature(model_id: str, feature_name: str, value: float) -> None: Adds or updates a feature for a specific model.
- get_features(model_id: str) -> Dict[str, float]: Retrieves the current features for a specified model.
- aggregate_features() -> Dict[str, Dict[str, float]]: Aggregates features for all models and returns them.Example 1:
Input: add_feature('gpt-3', 'latency', 200)
Output: NoneExample 2:
Input: get_features('gpt-3')
Output: {'latency': 200}Constraints:
- Up to 10,000 models can be processed at any time.

As OpenAI releases various models and updates, a systematic way to manage versions is necessary.
Problem statement: Design a class ModelVersionRegistry to manage different versions of AI models and associated metadata.
- add_version(model_id: str, version_id: str, metadata: dict) -> None: adds a new model version with its metadata.
- get_version(model_id: str, version_id: str) -> dict: retrieves the metadata for a specific model version.
- list_versions(model_id: str) -> List[str]: lists all versions of a given model.Example 1:
Input: add_version('gpt-3', 'v1', {'accuracy': 0.9})
Output: NoneExample 2:
Input: get_version('gpt-3', 'v1')
Output: {'accuracy': 0.9}Constraints:
- Each model may have up to 100 versions.

OpenAI needs efficient ways to calculate performance metrics of its AI models across large datasets.
Problem statement: Given a table model_performance with columns model_id, accuracy, and timestamp, write a SQL query to find the top 5 models by average accuracy over time, ordered from highest to lowest average accuracy.Example 1:
Input: N/A
Output: "SELECT model_id, AVG(accuracy) AS average_accuracy FROM model_performance GROUP BY model_id ORDER BY average_accuracy DESC LIMIT 5;"
Explanation: This query calculates the average accuracy of each model and retrieves the top 5.Example 2:
Input: N/A
Output: "SELECT model_id, AVG(accuracy) AS average_accuracy FROM model_performance WHERE timestamp > '2023-01-01' GROUP BY model_id ORDER BY average_accuracy DESC LIMIT 5;"
Explanation: This further restricts to models launched after a specific date.Constraints:
- model_performance contains at least 1 and at most 100,000 rows.

The ability to summarize large texts efficiently is crucial for OpenAI's API, allowing users to quickly digest information.
Problem statement: Given an input string text representing a large text body, your task is to implement a function summarize_text(text: str) -> str that uses a sliding window approach to summarize the text by extracting key sentences while maintaining coherence.
- summarize_text returns a str representing the summarized text.Example 1:
Input: "OpenAI provides a platform for developers to build applications using AI. The company focuses on promoting and advancing artificial intelligence. Their tools are user-friendly and accessible."
Output: "OpenAI provides a platform for developers to build applications using AI. Their tools are user-friendly and accessible."
Explanation: The function extracts 2 key sentences from the input text for the summary.Example 2:
Input: "Artificial intelligence has numerous applications, including natural language processing and robotics. OpenAI is at the forefront of this technology, constantly innovating products that leverage AI."
Output: "OpenAI is at the forefront of this technology, constantly innovating products that leverage AI."Constraints:
- text consists of 1 to 10^6 characters.
- The number of sentences in text will not exceed 100.