1Z0-1127-25 Practice Questions

T-Few is a Parameter-Efficient Fine-Tuning (PEFT) method. The core principle of PEFT techniques, including T-Few, is to freeze the vast majority of the pre-trained model's weights and update only a small, strategically chosen subset of parameters. This approach drastically reduces the number of trainable parameters, leading to significant reductions in computational requirements (GPU memory, processing time) and storage costs. Furthermore, by constraining the updates to a small parameter space, it helps mitigate catastrophic forgetting and prevents the model from overfitting to the smaller, specialized fine-tuning dataset, thereby improving generalization.

A. It updates all the weights of the model uniformly.

This describes standard or "vanilla" fine-tuning, not a parameter-efficient method like T-Few, which is defined by its selective updating of weights.

B. It selectively updates only a fraction of weights to reduce the number of parameters.

While it does update a fraction of weights, the primary goal isn't to reduce the model's total parameters but to reduce the trainable parameters for efficiency and better performance, as stated in C.

D. It increases the training time as compared to Vanilla fine-tuning.

This is incorrect. A key advantage of T-Few and other PEFT methods is that they are significantly faster and more computationally efficient than vanilla (full) fine-tuning.

---

1. Liu, H., Tam, D., Muqeeth, M., et al. (2022). Few-Shot Parameter-Efficient Fine-Tuning is Better and Cheaper than In-Context Learning. Advances in Neural Information Processing Systems (NeurIPS) 35. arXiv:2205.05638.

Section 1 (Introduction): The paper introduces T-Few as a method to address the high "training cost" of standard fine-tuning. It states, "we propose a set of parameter-efficient fine-tuning techniques... that is both more effective and dramatically more efficient." This directly supports the "reduce computational load" aspect of the correct answer.

2. Lialin, V., Deshpande, V., & Rumshisky, A. (2023). Scaling Down to Scale Up: A Guide to Parameter-Efficient Fine-Tuning. Journal of Machine Learning Research. arXiv:2303.15647.

Section 2 (Why PEFT?): This survey paper explicitly lists the motivations for PEFT methods. It states, "PEFT methods... offer a number of benefits over full fine-tuning," including "Reduces memory usage," "Reduces storage costs," and "Helps with catastrophic forgetting and overfitting." This directly validates the reasoning in option C.

3. Oracle Cloud Infrastructure (OCI) Documentation. Generative AI > Fine-tuning.

Section: Overview of Fine-Tuning: The official documentation explains that fine-tuning adapts a pre-trained model to a specific, smaller dataset. The principles discussed align with the goals of PEFT, which is to achieve this adaptation efficiently without the prohibitive cost of full fine-tuning and to improve performance on specific tasks, where avoiding overfitting is critical.

The Retrieval-Augmented Generation (RAG) Sequence model operates by first using a retriever component to find a fixed set of relevant documents (top-K) from a knowledge base based on the entire input query. This set of documents is then passed, along with the original query, to a generator model (like BART or T5). The generator conditions its entire output sequence on this same set of documents, effectively treating them as a latent variable and marginalizing over them to produce a single, cohesive, and contextually-grounded response. This approach ensures the generated text is informed by the retrieved knowledge for its full duration.

A. The RAG-Sequence model retrieves a set of top-K documents, not a single one, to provide a richer and more robust context for the generator.

C. The retriever component processes the entire input query to capture the full user intent, not just an initial segment, ensuring the most relevant documents are found.

D. The standard RAG-Sequence model uses the original input query for retrieval. Query modification is a different, more advanced technique not inherent to the base RAG-Sequence architecture.

1. Lewis, P., Perez, E., Piktus, A., Petroni, F., Karpukhin, V., Goyal, N., Küttler, H., Lewis, M., Yih, W., Rocktäschel, T., Riedel, S., & Kiela, D. (2020). Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks. In Advances in Neural Information Processing Systems 33 (NeurIPS 2020). Section 2.1, "Models," describes the RAG-Sequence model: "The RAG-Sequence model approximates the marginal by treating the top-K documents as a latent variable, z, and summing over the sequence probabilities, conditioned on each document..." This confirms a set of documents is used together.

2. Oracle Cloud Infrastructure Documentation. (2024). Overview of Generative AI. Section: "Retrieval Augmented Generation (RAG)." The documentation explains the general RAG pattern used in OCI Generative AI, stating, "The RAG model takes the user prompt and retrieves relevant information from your enterprise data... The prompt and the retrieved text chunks are presented as context to the LLM..." This supports the concept of retrieving a set of information (chunks/documents) to generate a response.

3. Stanford University. (2023). CS224N: NLP with Deep Learning, Lecture 15: Open-domain Question Answering. Stanford Center for Professional Development. The lecture notes describe RAG as a system where "Given a question, we first use a retriever... to retrieve a few relevant text passages from a large corpus... Then, we use a reader... to read the retrieved passages and output an answer." This aligns with retrieving a set of documents ("a few relevant text passages") for a given query.

String prompt templates are designed to be flexible constructs for generating dynamic prompts. They function by defining a base string with placeholders (variables) that are filled in at runtime. This design inherently supports any number of variables. A template can be a simple static string with zero variables, a string with a single placeholder, or a complex structure with multiple placeholders to accommodate various inputs, instructions, and contextual data.

A. Templates are not limited to a single variable; they are designed to handle multiple placeholders for complex, context-rich prompts.

B. The fundamental purpose of a prompt template is to incorporate variables to make a static string dynamic and reusable.

D. There is no minimum requirement for variables. A template with zero variables (a static prompt) or one variable is perfectly valid.

1. Oracle Cloud Infrastructure Documentation - Generative AI, Prompt Engineering: The official documentation describes prompts as containing "instructions, questions, input data, and examples." The "input data" component is variable, and the documentation places no restrictions on the number of such inputs, implicitly supporting zero to many.

Source: Oracle Cloud Infrastructure Documentation, "Generative AI > Prompt Engineering," Section: "Prompts."

2. Oracle Cloud Infrastructure Blog - Building with OCI Generative AI: Official tutorials and blog posts from Oracle frequently demonstrate the use of prompt templates with frameworks like LangChain. These examples explicitly show templates being created with multiple variables (e.g., {context} and {question}) as well as single variables, confirming the flexibility to use any number.

Source: Oracle Blogs, "Build a RAG solution with OCI Generative AI, LangChain, and OCI OpenSearch," (May 22, 2024), Code examples section.

3. OCI SDK for Python Documentation - Generative AI Examples: The code examples for the OCI Generative AI service often use Python's formatted strings (f-strings) to construct prompts. This method allows for the embedding of zero, one, or multiple variables directly into the prompt string, which serves as a practical implementation of a prompt template.

Source: Oracle Cloud Infrastructure SDK for Python Documentation, oci.generativeai.GenerativeAiClient, examples for generatetext.

In a typical LangChain run involving memory, the process follows a distinct two-step interaction. First, after the user provides input but before the main chain logic (e.g., the LLM call) executes, the chain reads from the memory object to load the conversational history. This history is used to construct the full context for the prompt. Second, after the core logic has executed and generated an output, the chain writes the latest user input and the model's response back into memory. This save operation occurs before the final output is returned to the user, ensuring the context is updated for subsequent interactions.

A. This is incorrect because memory is also read before the core logic is executed to provide the necessary context for the current turn.

B. This is incorrect because memory is accessed after receiving the current user input to include it in the context, not before.

D. This is incorrect because memory interaction is not a continuous stream but occurs at specific, discrete points in the execution cycle—primarily for loading and saving context.

1. LangChain Official Documentation, "How-to guides -> Chains -> Add memory to an LLMChain": This document explicitly outlines the interaction flow. In the section "Using memory in a chain," it states: "The Chain will use the memory in two ways. 1. First, it will load the memory variables and add them to the chain inputs... 2. After the chain has been executed, it will save the new input and output to memory." This directly supports the two-step process of reading before and writing after the core execution.

2. LangChain Official Documentation, "Conceptual guides -> Memory -> Getting started": This guide explains the fundamental purpose of memory is to "remember previous interactions with the user." This implies a cycle of loading past interactions to inform the present and saving the present interaction for the future, which aligns with the sequence described in the correct answer.

Full fine-tuning involves updating all the weights of a pre-trained model using a labeled, task-specific dataset. This process is computationally intensive and creates a full copy of the model for each task. In contrast, Parameter-Efficient Fine-Tuning (PEFT) methods, such as LoRA (Low-Rank Adaptation), freeze the original model's weights and only train a small number of new or existing parameters. This approach significantly reduces computational and storage costs while still effectively adapting the model to a new task using the same kind of labeled, task-specific data.

A. Continuous pretraining uses unlabeled, domain-specific data to adapt the model's general knowledge, not labeled, task-specific data.

B. Parameter-Efficient Fine-Tuning and Soft Prompting modify only a small subset of parameters (or add new ones) and use labeled data, not all parameters with unlabeled data.

D. Continuous pretraining modifies all of the model's original parameters, contrary to the claim that it requires no modification.

1. Oracle Cloud Infrastructure (OCI) Documentation, Generative AI, "Custom models and fine-tuning": "Fine-tuning updates all of the weights of a pretrained base model... This method is also known as full fine-tuning." and "Parameter-Efficient Fine-Tuning (PEFT) is a method that updates only a small subset of the model's parameters... PEFT is faster and less computationally expensive than full fine-tuning." This source directly supports the distinction made in option C.

2. Lialin, V., Deshpande, V., & Rumshisky, A. (2023). Scaling Down to Scale Up: A Guide to Parameter-Efficient Fine-Tuning. Section 2, "Background". arXiv:2303.15647. This paper defines full fine-tuning as updating all parameters of a pre-trained model for a downstream task. It then introduces PEFT methods as an alternative that tunes "only a small number of (extra) model parameters," which aligns perfectly with the description in option C.

3. Hu, E. J., Shen, Y., Wallis, P., Allen-Zhu, Z., Li, Y., Wang, S., ... & Chen, W. (2021). LoRA: Low-Rank Adaptation of Large Language Models. Section 3, "Our Method". arXiv:2106.09685. The foundational paper for LoRA, a popular PEFT technique, states: "We propose Low-Rank Adaptation, or LoRA, which freezes the pre-trained model weights and injects trainable rank decomposition matrices into each layer of the Transformer architecture." This confirms that PEFT methods add and train a few new parameters while keeping the original model frozen.

In the context of Large Language Models (LLMs), "hallucination" refers to the model generating content that is factually incorrect, nonsensical, or not grounded in the provided source data, yet presented with a high degree of confidence as if it were true. This phenomenon occurs because LLMs are probabilistic systems designed to predict the next most likely sequence of words based on their training data, rather than accessing a knowledge base to verify facts. As a result, they can invent details, fabricate sources, or create plausible-sounding but entirely false information. Mitigating hallucinations is a primary challenge in ensuring the reliability and trustworthiness of generative AI applications.

A. This describes creativity. While related, "hallucination" specifically refers to the undesirable generation of falsehoods, not desired imaginative output.

B. This is incorrect. Hallucination is a significant problem to be mitigated, not a technique used to improve model performance.

C. This describes a capability of multimodal models (e.g., image-to-text generation), not the definition of the term "hallucination."

1. Oracle Cloud Infrastructure (OCI) Documentation: In the official OCI Generative AI service FAQs, Oracle defines the term directly: "What is a hallucination? A hallucination is when a large language model (LLM) generates text that is nonsensical or not factually correct. Mitigating hallucinations is a key challenge in the field of generative AI."

Source: Oracle Cloud Infrastructure Documentation, "Generative AI - FAQs," Section: "What is a hallucination?".

2. Academic Publication: A comprehensive survey on the topic defines the phenomenon as follows: "Despite their remarkable capabilities, LLMs are prone to generating content that is nonsensical, unfaithful to the provided source, or otherwise untruthful. This phenomenon is commonly referred to as hallucination."

Source: Ji, Z., Lee, N., Frieske, R., Yu, T., Su, D., Xu, Y., ... & Fung, P. (2023). Survey of Hallucination in Natural Language Generation. ACM Computing Surveys, 55(12), Article 357, p. 2. DOI: https://doi.org/10.1145/3571730

3. University Courseware: Stanford University's Human-Centered AI Institute (HAI) explains the concept in its foundational materials: "When a large language model (LLM) generates text that is nonsensical or untethered from reality, it is said to be hallucinating... The model is simply generating a statistically likely sequence of words, and sometimes that sequence doesn’t align with our reality."

Source: Stanford University Human-Centered AI (HAI), "How to Tame AI Hallucinations," Published May 16, 2023.

OCI Dedicated AI Clusters are designed as a single-tenant, private supercluster for a customer's exclusive use. The defining characteristic is the complete isolation of the compute resources. This means the NVIDIA GPUs, the high-performance RDMA cluster network, and storage are not shared with any other customers. This isolation is crucial for generative AI workloads as it guarantees predictable, bare-metal performance without the "noisy neighbor" effect common in multi-tenant environments. It also provides a higher level of security for sensitive training data and proprietary models.

A. This describes a multi-tenant architecture, which is the opposite of a "Dedicated" cluster designed for single-tenant, exclusive use.

C. GPUs are specialized processors for massively parallel computation, which is their primary function in AI training, not data storage.

D. Dedicated AI Clusters use a private, ultra-low-latency RDMA over Converged Ethernet (RoCE) network, not the public internet, for inter-GPU communication.

1. Oracle Cloud Infrastructure Documentation, "Generative AI - Dedicated AI Clusters": "Dedicated AI clusters are for your exclusive use. You can create a dedicated AI cluster to fine-tune custom models or to host dedicated endpoints for your custom models." This statement directly confirms the single-tenant, isolated nature of the clusters. (Source: Oracle Cloud Infrastructure Official Documentation, Generative AI section).

2. Oracle Cloud Infrastructure Documentation, "AI Infrastructure - Overview": "OCI Supercluster is built using OCI Compute bare metal instances... The backend network is a high-performance RoCE v2 cluster network... This architecture provides a flat, non-blocking network at a massive scale, which is ideal for parallel workloads." This reference highlights the private, high-performance networking fabric, refuting the use of the public internet (Option D). (Source: Oracle Cloud Infrastructure Official Documentation, AI and Machine Learning -> AI Infrastructure section).

3. Oracle News, "Oracle Announces New OCI Supercluster to Bring the Power of AI to More Organizations" (May 23, 2023): "With OCI Supercluster, OCI is extending its AI infrastructure to give customers a dedicated, managed, and easy-to-deploy AI supercomputer for their most demanding AI workloads." The term "dedicated" is used to emphasize the single-customer, isolated environment. (Source: Oracle Official Press Release).

Vector databases are fundamentally architected to manage and query high-dimensional data, known as vector embeddings. Unlike the tabular structure of relational databases, their core operational principle is based on measuring the proximity between vectors. They use specialized indexing algorithms (e.g., HNSW, IVF) to efficiently perform similarity searches by calculating distances (like Euclidean or Manhattan) or similarity scores (like cosine similarity) in a multi-dimensional vector space. This allows them to find the "nearest neighbors" to a given query vector, which is essential for applications like semantic search, recommendation systems, and Retrieval-Augmented Generation (RAG).

A. It stores data in a linear or tabular format.

This describes the structure of traditional relational databases, which organize data into rows and columns within tables, not the multi-dimensional structure of vector databases.

B. It is not optimized for high-dimensional spaces.

This is incorrect. The primary advantage and purpose of a vector database is its optimization for efficient storage and querying of data in high-dimensional spaces.

C. It uses simple row-based data storage.

This is characteristic of relational databases. Vector databases employ complex, specialized data structures and indexing mechanisms designed for proximity searches, not simple row-based storage.

1. Oracle Cloud Infrastructure Documentation, "AI Vector Search Overview." This document states, "OCI AI Vector Search is a fully managed service that enables you to search for similar vectors in a high-dimensional space... You can then query the index to find the vectors that are most similar to your query vector." This directly supports that the structure is based on similarity.

2. Oracle Database 23ai Documentation, "AI Vector Search Developer's Guide," Section 1.1, "Introduction to AI Vector Search." The guide explains, "You can then query this data by finding the vectors that are closest to your query vector. This is called a similarity search or a vector search." This confirms the operational basis is distance and similarity.

3. Stanford University, CS224N: NLP with Deep Learning, Winter 2023, Lecture 2 Notes, "Word Vectors and Word Senses," Section 2. The courseware details how word vectors (embeddings) are positioned in a vector space where distance represents semantic similarity. It notes, "We can also use the cosine similarity of two vectors... as a metric of how similar the words are." This foundational concept underpins the architecture of vector databases.

Fine-tuning is the appropriate method when a pre-trained Large Language Model (LLM) does not perform well on a specific, nuanced task, and the number of examples needed to guide it via prompt engineering (in-context learning) is too large to fit within the context window or is inefficient. Fine-tuning adapts the model's internal weights by training it on a curated dataset of examples, effectively "teaching" it the new skill, style, or format. This process internalizes the knowledge, leading to better performance and consistency on the target task compared to what can be achieved with prompting alone, especially when the task is complex.

A. If the LLM already understands the necessary topics, prompt engineering is typically a more efficient and cost-effective method for guiding its output.

C. Providing an LLM with access to the latest data is the primary use case for Retrieval-Augmented Generation (RAG), which retrieves current information and adds it to the prompt's context.

D. Fine-tuning is a supervised learning process that explicitly requires a structured dataset of instructions and examples to optimize the model for a specific task.

1. Oracle Cloud Infrastructure (OCI) Documentation: In the documentation for the Generative AI service, it specifies when to use custom models. "Fine-tuning can improve on few-shot learning by training the model on many more examples than can fit in a prompt, making fine-tuning a good option for new tasks that the model wasn't originally trained on."

Source: Oracle Cloud Infrastructure Documentation, "AI and Machine Learning > Generative AI > Custom Models", Section: "When to Use a Custom Model".

2. Stanford University Courseware: Lecture materials on LLM adaptation explain that fine-tuning is used to specialize a model for a target task or data distribution that is not well-handled by the base model, especially when in-context learning is insufficient.

Source: Stanford CS224N: NLP with Deep Learning, Winter 2023, Lecture 12: "Prompting, Instruction Finetuning", Slide 54, "Why instruction finetuning?".

3. Academic Publication: A foundational paper on instruction tuning highlights that this form of fine-tuning significantly improves zero-shot performance on unseen tasks, demonstrating its value when a model needs to learn how to perform new tasks it struggles with initially.

Source: Wei, J., Bosma, M., et al. (2021). "Finetuned Language Models Are Zero-Shot Learners". Section 3.1, "Method". (Available on arXiv:2109.01652). This paper discusses the principle of fine-tuning to improve task performance where the base model is lacking.

Prompt templates are structured, reusable text formats that serve as a blueprint or "recipe" for generating prompts. They contain placeholders (variables) that are filled in with specific data or context before being sent to a language model. This approach ensures consistency, guides the model's response structure, and simplifies the process of creating effective prompts for various tasks. By defining the context, instructions, and input data format, templates guide the model to produce more predictable and relevant outputs.

A. Prompt templates are text-based structures with placeholders, not complex algorithms that require compilation like source code.

C. The fundamental purpose of a template is to be modified and customized by filling in variables with specific context or data.

D. Prompt templates are designed primarily for textual content, as this is the native input format for language models.

1. Official Vendor Documentation: Oracle Cloud Infrastructure (OCI) Documentation, "Generative AI - Prompt Engineering". The documentation explains that a prompt consists of instructions and context, which can be structured for reuse. This structured format with placeholders is functionally a template or recipe. The guide states, "A good prompt gives the model enough context... You can also provide few-shot examples in the prompt to guide the model." This practice of creating a structured, reusable format is the essence of a template. (See OCI Documentation > AI and Machine Learning > Generative AI > Key Concepts > Prompt Engineering).

2. Academic Publication: Liu, P., Yuan, W., Fu, J., et al. (2023). "Pre-train, Prompt, and Predict: A Systematic Survey of Prompting Methods in Natural Language Processing." ACM Computing Surveys, 55(9), Article 195, p. 6. The paper discusses "Template Construction," describing templates as textual strings containing slots (placeholders) for input text. It states, "The first step of prompt-based learning is to design a template, which is a textual string that has two slots...". This definition aligns directly with the concept of a predefined recipe. DOI: https://doi.org/10.1145/3560815

3. University Courseware: Stanford University, CS324 - Large Language Models, Winter 2022, "Lecture 3: Prompting, Instruction Finetuning". The course materials discuss how prompts are structured to elicit desired behavior. The concept of creating a format (e.g., "Translate English to French: {englishtext} =>") is presented, which is a form of templating or a recipe for a specific task.