LLM Quiz - Questions & Answers
What practice would help reduce hallucinations in an LLM giving factual advice?
-
Using open-ended questions
-
Providing specific source references in prompts ✅
-
Allowing the model to guess the missing data
-
Encouraging speculative responses
Which of the following strategies is least effective in reducing hallucinations in language models?
-
Reinforcement learning from human feedback (RLHF)
-
Using a smaller dataset for training ✅
-
Fine-tuning on domain-specific data
-
Incorporating factual consistency checks
Which of the following statements is true about the licensing of open-source LLMs?
-
Open-source LLMs cannot be used for commercial purposes
-
All open-source LLMs must be distributed under the Apache License
-
Open-source LLMs can have a variety of licenses, some of which may impose specific usage restrictions ✅
-
Open-source LLMs always come with no usage restrictions
How does a high temperature value affect the probability distribution of the next token in LLM outputs?
-
It sharpens the distribution, making high-probability tokens more likely
-
It flattens the distribution, making low-probability tokens more likely ✅
-
It has no effect on the distribution.
-
It biases the model towards shorter responses
Which scenario best exemplifies the use of one-shot prompting?
-
Providing a detailed list of instructions followed by multiple examples
-
Giving one example of a complex task and expecting the model to generalize ✅
-
Using a large dataset to train the model incrementally
-
Setting up a reinforcement learning environment with numerous iterations
What is the primary issue with the "bias amplification" phenomenon in AI systems?
-
It causes AI models to underperform in terms of accuracy
-
It leads to the reinforcement and exaggeration of existing biases in the data ✅
-
It makes AI systems more sensitive to noise in the input data
-
It results in overfitting the training data
In one-shot prompting, the primary goal is to
-
Train a model from scratch with a single data point
-
Fine-tune a pre-trained model using a single example
-
Generate a desired response from a model with one example in the prompt ✅
-
Use multiple examples to improve model accuracy
In what ways can the efficacy of prompts in multilingual models be improved?
-
Applying language-specific nuances ✅
-
Using translation tools ✅
-
Avoiding cultural references
-
Providing examples in multiple languages ✅
In a customer recommendation system, how can hallucination errors be minimized?
-
Process real customer feedback in updates ✅
-
Ensure realistic constraints in prompts ✅
-
Provide probabilistic confidence scores ✅
-
Ignore rare and unique customer behavior patterns
What is a potential drawback of few-shot prompting that practitioners should be aware of?
-
High computational costs during inference
-
Lack of flexibility in adapting to new tasks
-
Risk of overfitting to the examples in the prompt ✅
-
Requirement of large-scale pre-training datasets
Which of the following is a key difference between the development communities of open-source and closed-source LLMs?
-
Open-source communities typically involve contributions from a wide range of independent developers and organizations ✅
-
Closed-source communities often have more diverse contributions from various stakeholders
-
Open-source communities do not allow any external contributions
-
Closed-source communities are known for having more transparent development processes
Which of the following strategies is most effective for reducing the length of responses generated by a language model without significantly compromising on the quality of the response?
-
Reducing the temperature parameter.
-
Using more specific and detailed prompts
-
Setting a maximum token limit for the response ✅
-
Increasing the batch size during inference
In the context of preventing hallucinations in generative AI models, what does "model distillation" refer to?
-
Reducing the model size by approximating a larger model ✅
-
Using distilled water to cool down AI hardware
-
Training a model on distilled facts to ensure accuracy
-
Distilling the model's training process to essential components only
In the context of fine-tuning an LLM for a specific application, why might one opt to use a lower temperature setting during inference?
-
To encourage the generation of highly diverse outputs
-
To enhance the randomness and creativity of the model
-
To reduce hallucinations and generate more precise responses ✅
-
To increase the likelihood of generating factual inaccuracies
What steps can be taken to ensure LLMs provide culturally sensitive outputs?
-
Curate culture-specific datasets with diverse perspectives ✅
-
Employ region-specific context in prompts ✅
-
Use a single culture's dataset to maintain consistency
-
Validate outputs by cultural experts ✅
A generative AI used for educational content sometimes includes outdated information. What methods can address this?
-
Regular updates with the latest academic research ✅
-
Encouraging model creativity over factual accuracy
-
Cross-verifying outputs with up-to-date references ✅
-
Incorporate a feedback mechanism for educators ✅
Which of the following scenarios would most benefit from using a higher temperature setting for an LLM?
-
Summarizing legal documents
-
Generating poetry or creative writing ✅
-
Answering factual questions
-
Translating technical manuals
When optimizing prompts for generating structured outputs (like JSON), which of the following modifications can significantly improve the model's accuracy in producing the desired structure?
-
Adding explicit instructions to the prompt. ✅
-
Training the model on a smaller dataset with similar structures
-
Using a higher learning rate during training
-
Reducing the model's context window
Which of the following approaches is best suited for optimizing prompts to ensure that a language model generates responses that are both concise and contextually relevant?
-
Increasing the model's attention span
-
Incorporating keywords related to the desired context
-
Iteratively refining the prompt based on generated outputs ✅
-
Increasing the number of training epochs
How can developers ensure generative AI avoids spreading misinformation?
-
Using current and reliable sources ✅
-
Regularly updating the model with new training data ✅
-
Implementing cross-referencing mechanisms within the model ✅
-
Encouraging creativity over accuracy