Scenario practice questions

Which technique allows an LLM to be adapted to a new task with only a few examples?

Trap 1: Fine-tuning

Fine-tuning requires more data and computational resources.

Trap 2: Pre-training

Pre-training trains from scratch on large data, not for few-shot adaptation.

Trap 3: Prompt engineering

Prompt engineering is broader; few-shot is a subset but the question asks for a specific technique.

A
Few-shot learning
Few-shot learning provides examples in the prompt to adapt the model to a new task.
B
Fine-tuning
Why wrong: Fine-tuning requires more data and computational resources.
C
Pre-training
Why wrong: Pre-training trains from scratch on large data, not for few-shot adaptation.
D
Prompt engineering
Why wrong: Prompt engineering is broader; few-shot is a subset but the question asks for a specific technique.

Question 2hardmulti select

A troubleshooting scenario: A RAG system returns no results for certain queries. The index exists and has documents. Which TWO are likely causes?

Trap 1: The search algorithm is set to exact (brute-force) and index is…

Exact search still finds matches if they exist.

Trap 2: The query is too long for the embedding model

Models typically truncate long inputs, but still produce an embedding.

Trap 3: The index has not been refreshed after adding new documents

New documents are immediately searchable unless refresh interval is disabled.

A
The search algorithm is set to exact (brute-force) and index is small
Why wrong: Exact search still finds matches if they exist.
B
The query embedding dimension does not match the index dimension
Dimension mismatch causes search errors or zero results.
C
The query is too long for the embedding model
Why wrong: Models typically truncate long inputs, but still produce an embedding.
D
The index has not been refreshed after adding new documents
Why wrong: New documents are immediately searchable unless refresh interval is disabled.
E
The embedding model used for indexing is different from the one used for query
Different models produce incompatible embeddings, so no results match.

Question 3hardmultiple choice

A company is deploying a multi-language chatbot using OCI Generative AI Service. The chatbot must support English, Spanish, and French. The team finds that responses in Spanish are less accurate than in English. They have a small bilingual dataset. What is the best approach?

Trap 1: Use prompt engineering with language-specific instructions in the…

Prompt engineering may not be enough if the base model has limited Spanish capabilities.

Trap 2: Translate all user queries to English, process them, then translate…

This adds latency and potential translation errors; it's not optimal for real-time chatbots.

Trap 3: Train separate fine-tuned models for each language.

This is resource-intensive and not necessary; a multilingual model can handle all languages.

A
Use a multilingual base model (e.g., mT5) and fine-tune on the bilingual dataset (English and Spanish) using cross-lingual transfer learning.
Cross-lingual transfer leverages English data to improve Spanish performance, and fine-tuning on bilingual data further boosts accuracy.
B
Use prompt engineering with language-specific instructions in the system prompt.
Why wrong: Prompt engineering may not be enough if the base model has limited Spanish capabilities.
C
Translate all user queries to English, process them, then translate responses back.
Why wrong: This adds latency and potential translation errors; it's not optimal for real-time chatbots.
D
Train separate fine-tuned models for each language.
Why wrong: This is resource-intensive and not necessary; a multilingual model can handle all languages.

Question 4mediummultiple choice

A data scientist deployed a fine-tuned Llama 2 7B model on OCI Model Deployment with a single VM.GPU.A10.1 shape. Users report average latency of 3 seconds per request, which is too high for the intended real-time application. The model is used for short text generation (max 128 tokens). The data scientist wants to reduce per-request latency without significant accuracy loss. Which action would be most effective?

Trap 1: Increase the number of workers per replica

More workers on a single GPU can lead to contention and may not reduce per-request latency.

Trap 2: Increase the max_tokens parameter for the model

Increasing max_tokens increases the output length, thereby increasing latency.

Trap 3: Enable response streaming for the model endpoint

Streaming allows partial output but does not reduce the total processing time.

A
Increase the number of workers per replica
Why wrong: More workers on a single GPU can lead to contention and may not reduce per-request latency.
B
Increase the max_tokens parameter for the model
Why wrong: Increasing max_tokens increases the output length, thereby increasing latency.
C
Enable response streaming for the model endpoint
Why wrong: Streaming allows partial output but does not reduce the total processing time.
D
Apply 4-bit quantization using AWQ
Quantization reduces model size and inference time with minimal accuracy loss.

Question 5hardmultiple choice

A company is using OCI GenAI with a Dedicated AI Cluster to serve a large language model for real-time chat applications. They notice high inference latency (average 2 seconds per response) and want to reduce it to under 500 milliseconds without significantly degrading the quality of responses. The cluster is configured with NVIDIA A100 GPUs. The model is the base Cohere Command model (52B parameters). They have explored increasing batch size, but that increases latency for interactive use cases. Which action should they take?

Trap 1: Increase batch size to process multiple queries at once.

Batch processing increases latency for each individual query in real-time scenarios.

Trap 2: Swap the model to a smaller variant, such as Cohere Command Light…

Smaller models have lower quality, which the team wants to avoid.

Trap 3: Enable model quantization (e.g., int8) to reduce memory and…

Quantization can degrade output quality, which is not acceptable.

A
Deploy the model with inference optimization frameworks like vLLM, TensorRT, or ONNX Runtime.
These frameworks optimize GPU utilization and reduce latency without changing the model.
B
Increase batch size to process multiple queries at once.
Why wrong: Batch processing increases latency for each individual query in real-time scenarios.
C
Swap the model to a smaller variant, such as Cohere Command Light (6B).
Why wrong: Smaller models have lower quality, which the team wants to avoid.
D
Enable model quantization (e.g., int8) to reduce memory and computation.
Why wrong: Quantization can degrade output quality, which is not acceptable.

Question 6mediummultiple choice

A financial institution uses OCI GenAI to power a customer support chatbot. The compliance team requires that responses are strictly consistent with regulatory guidelines and approved responses. The company has a curated set of question-answer pairs that cover common scenarios. They want to ensure that the chatbot never deviates from these approved answers. The data science team is considering various approaches to enforce this consistency. Which approach is most effective?

Trap 1: Few-shot prompting with three example responses in every query.

Few-shot examples may not cover all variations, and the model may still generate unapproved content.

Trap 2: Using a large context window to include all regulatory guidelines…

Even with a large context, the model may not consistently follow the guidelines.

Trap 3: Setting a low temperature (0.1) to make outputs deterministic.

Low temperature reduces randomness but does not constrain content to approved responses.

A
Few-shot prompting with three example responses in every query.
Why wrong: Few-shot examples may not cover all variations, and the model may still generate unapproved content.
B
Fine-tuning the model on the curated dataset of question-answer pairs.
Fine-tuning adapts the model to mimic the approved responses, providing strong consistency.
C
Using a large context window to include all regulatory guidelines in the prompt.
Why wrong: Even with a large context, the model may not consistently follow the guidelines.
D
Setting a low temperature (0.1) to make outputs deterministic.
Why wrong: Low temperature reduces randomness but does not constrain content to approved responses.