The term "Microgpt" does not refer to a standardized product with universally defined input data formats; rather, it typically describes a small-scale, custom-built language model or an application interfacing with such a model. Therefore, the data formats accepted depend entirely on the specific implementation, the libraries used for its development (e.g., PyTorch, TensorFlow, Hugging Face Transformers) ， and its intended purpose. However, for any language model, whether for training or inference, the fundamental input format is text. For training, this text can be structured in various common data formats like plain text files, JSON, JSON Lines, or CSV. For inference, the input is most commonly a plain string prompt, although structured JSON might be used for specific API interactions.
For the purpose of training a "Microgpt," developers typically prepare datasets in formats that allow efficient loading and processing. Plain text files (.txt) are often used for large bodies of unstructured text, where each file or distinct block of text can serve as a training example. For more structured data, JSON (.json) or JSON Lines (.jsonl) are preferred, especially when each text sample comes with associated metadata (e.g., {"id": "doc1", "text": "This is a document.", "category": "news"}) . JSON Lines, where each line is a valid JSON object, is particularly useful for handling very large datasets that can be processed line by line. CSV (.csv) or TSV (.tsv) files are also common for tabular data where one or more columns contain the text to be processed, alongside other features. Regardless of the initial format, all raw text data undergoes preprocessing steps such as tokenization, numericalization (converting tokens to numerical IDs) ， padding, and creation of attention masks before being fed into the model's neural network.
When interacting with a "Microgpt" for inference, the most straightforward input is usually a plain string representing a query or prompt, such as "Generate a summary of natural language processing." If the "Microgpt" is part of a larger application or accessed via an API, structured inputs like JSON objects might be used to provide the prompt along with other parameters, for example, {"prompt": "Explain vector embeddings.", "max_tokens": 100}. In many advanced applications, especially those leveraging retrieval-augmented generation (RAG) ， the input to the "Microgpt" is often enriched with external context. This context is frequently retrieved from a vector database. A user's text query is first converted into a vector embedding, which is then used to perform a similarity search against a collection of vectorized text segments stored in a vector database like Zilliz Cloud . The retrieved text segments, originally stored in formats like plain text or JSON before being embedded, are then appended to the original query as additional context, allowing the "Microgpt" to generate more informed and relevant responses.