Generative AI and LLM glossary

Models

Sarvam 105B is Sarvam AI's flagship open-weight 105B+ MoE reasoning model for Indian-language chat, coding, long-context work, and agents.

AI platforms

Sarvam AI is an India-based sovereign AI platform focused on Indian-language LLMs, speech, translation, document digitization, and enterprise AI agents.

Machine Learning

A learning approach that combines labeled and unlabeled data for training models.

Models

Speech-to-speech, or full-duplex voice, takes spoken audio and returns spoken audio in one live session instead of chaining STT, an LLM, and TTS.

Development

Structured Outputs constrain model responses to match a developer-supplied schema, typically for reliable JSON or function-call arguments.

Data Science

The process of creating artificial data that mimics real-world data for training machine learning models.

Models

A time-series foundation model is pretrained on many series so it can forecast new series with little or no task-specific training.

Machine Learning

A type of machine learning that deals with data that has no labels, aiming to find hidden patterns or intrinsic structures.

Model Evaluation

Variance measures a model's sensitivity to fluctuations in the training data, contributing to overfitting when high.

Retrieval

In AI search, a vector is a list of numbers that locates an item in embedding space so nearby vectors can be retrieved as similar.

Machine Learning

A machine learning approach where the model predicts classes that it has not seen during training.