What is a "token" in the context of how a Large Language Model processes text?

A small chunk of text, roughly a word or part of a word. An LLM first splits text into tokens. A token is roughly a word or part of a word. Research on subword units showed that a rare word can be written as several smaller pieces. The model then calculates a probability for every possible next token. The other options describe a security key, GPU memory and a training example. None of these is a token. Read more: SciTech0136 — Large Language Models (LLMs).

Modern Large Language Models are built on an architecture introduced in the 2017 paper "Attention Is All You Need". What is this architecture called?

Transformer. The 2017 paper proposed the Transformer. It is based only on attention and drops recurrence and convolutions. It trains faster because many computations run in parallel. GPT-3 uses transformer layers, and the largest Llama 3 model is a dense Transformer. Recurrent and convolutional networks were the older approach that the paper replaced. Read more: SciTech0136 — Large Language Models (LLMs).

During training, an LLM's prediction error is measured as a loss. Which kind of method then adjusts its parameters to reduce that loss?

Gradient-based optimisation, such as the Adam optimiser. Pre-training measures error as a cross-entropy loss. A gradient-based optimiser then nudges the parameters to lower it. GPT-3 used the Adam optimiser and clipped the gradient at 1.0. The Adam paper describes Adam as an algorithm for first-order gradient-based optimization. Retrieval is a separate technique that helps at answer time, not in training. Read more: SciTech0136 — Large Language Models (LLMs).

Consider the following statements about Large Language Models: 1. Biases in training data can lead an LLM to produce stereotyped content. 2. Hallucination means producing fluent, confident text that is false. 3. Fine-tuning with human feedback removes hallucination completely. Which of the statements given above is/are correct?

1 and 2 only. Statements 1 and 2 are correct. The GPT-3 paper warns that biased data can give stereotyped output. NIST defines confabulation as confidently stated but false content. Statement 3 is wrong. InstructGPT improved truthfulness with human feedback, but its authors say it still makes simple mistakes. Read more: SciTech0136 — Large Language Models (LLMs).

A PIB release of 13 February 2026 listed support under the IndiaAI Innovation Centre (Foundation Models) pillar. Which organisation received the largest total support?

BharatGen, the IIT Bombay-led consortium. BharatGen received Rs 1,058.52 crore, made of Rs 990.92 crore compute support and Rs 67.60 crore other costs. Sarvam received Rs 246.72 crore. Gnani received Rs 177.27 crore and Soket AI Rs 177.08 crore. Twelve organisations were selected in all. PIB of 25 November 2025 says BharatGen is supported under the NM-ICPS of the Department of Science and Technology. It says Rs 235 crore is channelled through the Technology Innovation Hub at IIT Bombay. Read more: SciTech0136 — Large Language Models (LLMs).

The Cabinet approved the IndiaAI Mission on 7 March 2024. What budget outlay did it approve?

Rs 10,371.92 crore. The Cabinet approved Rs 10,371.92 crore for the IndiaAI Mission, to be run by the IndiaAI Independent Business Division under the Digital India Corporation. Rs 1,058.52 crore is BharatGen's support and Rs 246.72 crore is Sarvam's. Rs 76,000 crore is the Semicon 1.0 outlay, a separate semiconductor programme. Read more: SciTech0136 — Large Language Models (LLMs).

MCQ QuestionsMy progress

As of August 2026, from 506 applications, how many indigenous foundation model proposals had the Government selected under the IndiaAI Mission?

Question 1 of 7

Was this helpful?

As of August 2026, from 506 applications, how many indigenous foundation model proposals had the Government selected under the IndiaAI Mission?

Keep practicing

Beyond the answer

Leave a ReplyCancel reply

Discover more from MCQ Questions