Alibaba Cloud ACA Generative AI Engineer

Which architectural component introduced in the 2017 Vaswani et al paper is the foundation of modern large language models such as Alibaba's Qwen series?

• ARecurrent neural network (RNN)
• BTransformer with self-attention
• CLong short-term memory (LSTM)
• DConvolutional neural network (CNN)

In a Transformer self-attention block, what role do the Query, Key and Value projections play?

• AQuery and Key compute attention weights, which are then used to combine Values into the o…
• BQuery produces the output, Key and Value are discarded after training
• CQuery stores the vocabulary, Key stores positions, Value stores embeddings
• DAll three are identical copies of the input embedding

Which Transformer variant is best suited for autoregressive text generation as used by Qwen-Chat?

• AEncoder-only (BERT-style)
• BDecoder-only with causal masking
• CEncoder-decoder (T5-style)
• DBidirectional encoder with pooling

What does Byte Pair Encoding (BPE) do during LLM tokenization?

• AEncrypts user prompts before sending them to the model
• BIteratively merges the most frequent adjacent character or subword pairs to build a fixed…
• CSplits text strictly on whitespace and punctuation
• DConverts each Unicode character into its ASCII equivalent

An embedding model maps a sentence to a 1024-dimensional dense vector. Which downstream task is the most natural use for this vector?

• AStoring in a vector database for semantic similarity search
• BDecoding back into the original sentence character by character
• CReplacing the model's tokenizer at inference time
• DServing as a system prompt

Which statement best describes a foundation model in the context of generative AI?

• AA small task-specific classifier trained from scratch on labelled data
• BA large model pre-trained on broad data that can be adapted to many downstream tasks
• CA rule-based expert system with hand-crafted prompts
• DA model that only runs on edge devices

Scaling laws for LLMs (Kaplan, Chinchilla) predict that, holding compute fixed, what most improves test loss?

• AIncreasing batch size only
• BBalancing model parameters and training tokens roughly proportionally
• CAlways making the model larger regardless of data
• DReducing the number of layers

What is meant by 'emergent abilities' in large language models?

• ACapabilities that appear only at sufficient model scale and are not present in smaller mo…
• BSkills that emerge during fine-tuning but disappear at inference
• CHard-coded heuristics shipped by the model vendor
• DFeatures that only work in Chinese

Which sequence correctly describes the typical lifecycle of a chat-tuned LLM such as Qwen-Chat?

• AFine-tuning then pre-training then RLHF
• BPre-training then supervised fine-tuning then preference alignment
• CRLHF then BPE then pre-training
• DQuantization then pre-training then SFT

In RLHF, what is the role of the reward model?

• AIt generates the next token directly
• BIt assigns a scalar quality score to a response that the policy is then trained to maximi…
• CIt encrypts the policy weights
• DIt tokenizes user input

Direct Preference Optimization (DPO) is often used as an alternative to RLHF because it

• AEliminates the need for an explicit reward model and uses preference pairs directly with…
• BRequires more reward annotations than PPO
• CTrains a new tokenizer for each task
• DIs the only method that works for image generation

LoRA (Low-Rank Adaptation) reduces fine-tuning cost by

• AQuantizing the entire base model to 4 bits during training
• BInserting small low-rank matrices into selected weight projections and training only those
• CTraining a brand new model from scratch with fewer layers
• DRemoving the attention mechanism

QLoRA differs from standard LoRA primarily because it

• ALoads the base model in 4-bit NF4 quantization while training LoRA adapters in higher pre…
• BUses no adapters at all
• CRequires twice the GPU memory of LoRA
• DOnly works on encoder-only models

Prefix tuning is a parameter-efficient fine-tuning method in which

• AA small set of trainable virtual tokens (the prefix) are prepended to each layer's input…
• BAll Transformer layers are retrained from scratch
• CThe vocabulary is doubled
• DOnly the embedding layer is unfrozen

Within Alibaba's Tongyi Qianwen family, which model is the multimodal vision-language variant?

• AQwen-Coder
• BQwen-Audio
• CQwen-VL
• DQwen-Math

Which Qwen model size is most appropriate for cost-sensitive on-device or edge deployments where memory is constrained?

• AQwen-72B base
• BQwen2-1.5B or Qwen2-0.5B
• CQwen-Max via DashScope
• DQwen-14B-Chat

What is the difference between a Qwen 'base' model and a Qwen 'Chat' model?

• ABase is for next-token completion only; Chat has been instruction-tuned and aligned for d…
• BChat has fewer parameters than the base
• CBase supports more languages than Chat
• DChat is open weight; Base is closed source

DashScope is best described as

• AAlibaba Cloud's managed model-invocation API for Qwen and other Tongyi models
• BAn open-source vector database
• CA GPU rental service only
• DA replacement for Object Storage Service (OSS)

Why would you enable streaming mode (server-sent events) on a DashScope chat call?

• ATo return tokens incrementally so users see output as it is generated, reducing perceived…
• BTo increase the model's context window
• CTo skip safety filtering
• DTo compress the prompt

Function calling on DashScope lets a Qwen model

• AReturn a structured JSON tool-call request that the application can execute and then feed…
• BRun arbitrary shell commands inside the model server
• CReplace the system prompt at runtime
• DCompress responses to half their normal length

Which PAI sub-product is the managed elastic inference service used to deploy LLMs as REST/gRPC endpoints with autoscaling?

• APAI-DLC
• BPAI-EAS
• CPAI-DSW
• DPAI-Designer

PAI-DLC (Deep Learning Containers) is most useful when you need to

• ARun a distributed multi-node, multi-GPU training job for a Qwen fine-tune
• BEdit a notebook on a single CPU
• CServe a real-time chat endpoint
• DStore embeddings for similarity search

PAI-DSW (Data Science Workshop) primarily provides

• AA managed JupyterLab environment with pre-installed deep-learning frameworks and GPU acce…
• BA drag-and-drop pipeline builder for non-coders
• CA serverless model endpoint with autoscaling
• DA vector database

Which PAI feature is intended for users who want to build ML pipelines visually without writing code?

• APAI-DLC
• BPAI-EAS
• CPAI-Designer
• DPAI-Lingjun

PAI-Lingjun (Smart Computing Platform) is differentiated from generic GPU clusters by its

• AHigh-throughput RDMA interconnect optimised for large-scale GenAI training and inference
• BLimit to single-GPU workloads
• CSupport for CPU-only inference only
• DLack of any networking layer

PAI-AutoML is most appropriate when you want to

• AAutomatically search hyperparameters and architectures for a tabular or vision problem
• BManually set every learning rate by hand
• CSkip data preparation entirely
• DReplace your serving endpoint

Zero-shot prompting refers to

• AAsking the model to perform a task with only an instruction and no in-context examples
• BAlways providing exactly five worked examples
• CUsing a model that was never pre-trained
• DStreaming the prompt one word at a time

Few-shot prompting is most likely to outperform zero-shot when

• AThe task has a specific output format or label space the model must match
• BThe model has unlimited context length and any task
• CYou want to minimise prompt token cost above all else
• DYour task is creative free-form writing

Chain-of-thought (CoT) prompting improves performance on complex reasoning tasks because it

• AEncourages the model to produce intermediate reasoning steps before the final answer
• BDoubles the model parameters at runtime
• CReplaces the tokenizer
• DDisables the safety filter

Self-consistency improves CoT by

• ASampling multiple reasoning chains with temperature > 0 and majority-voting on the final…
• BForcing greedy decoding only
• CUsing a smaller model for the final answer
• DSkipping the reasoning steps

ReAct prompting interleaves

• AReasoning steps and tool actions, with each action's observation feeding the next thought
• BTwo different LLMs taking turns generating tokens
• CImage and audio modalities only
• DUser and system messages randomly

When designing a chat prompt with Qwen-Chat, what is the conventional purpose of the system message versus the user message?

• ASystem sets persistent role, tone and constraints; user supplies the per-turn request
• BUser sets the persona and system supplies the data
• CThey are interchangeable and the API ignores roles
• DSystem is reserved for safety filtering only

Role-playing prompts (e.g. 'You are an expert oncologist') tend to improve quality because they

• ASteer the model toward higher-quality regions of the training distribution associated wit…
• BIncrease the number of training tokens
• CSwitch the model to a different architecture
• DReduce safety guardrails

When picking few-shot exemplars, the most reliable practice is to

• AChoose examples that are diverse, representative of edge cases and similar in format to t…
• BPick the longest possible examples
• CAlways reuse the same single example
• DUse random untranslated text

Retrieval-Augmented Generation (RAG) primarily addresses which weakness of vanilla LLMs?

• ALack of access to up-to-date or proprietary knowledge beyond the training cut-off
• BInability to process Chinese characters
• CSlow tokenization
• DRandom GPU failures

Which RAG chunking strategy preserves semantic boundaries by splitting at paragraph or sentence breaks rather than fixed token counts?

• AFixed-size chunking
• BRecursive character chunking with separator hierarchy
• CRandom chunking
• DSingle-document chunking

Semantic chunking differs from recursive character chunking because it

• ASplits text where the embedding similarity between adjacent sentences drops below a thres…
• BUses random offsets
• CAlways produces chunks of exactly 512 tokens
• DIs faster than fixed-size chunking

Which embedding model family is widely used for Chinese-English RAG on Alibaba Cloud and is available through DashScope and ModelScope?

• Abge (BAAI General Embedding)
• BImageNet ResNet
• CMNIST CNN
• DAlexNet

Which Alibaba Cloud product provides PostgreSQL-compatible vector search via the pgvector extension and is often paired with Qwen for RAG?

• AHologres
• BOSS
• CECS
• DFunction Compute

Tair-Vector is a vector capability built on top of which Alibaba Cloud service?

• ATair (an enhanced Redis-compatible in-memory database)
• BMaxCompute
• CObject Storage Service (OSS)
• DPolarDB-X

AnalyticDB for PostgreSQL Vector edition is designed for

• ALarge-scale analytic workloads that combine SQL analytics with vector similarity search
• BReplacing the BPE tokenizer
• CStoring model weights
• DGPU scheduling

OpenSearch Vector Search Edition on Alibaba Cloud is most useful for RAG pipelines that need

• AHybrid retrieval combining BM25 keyword scoring and dense vector similarity
• BPure SQL OLTP transactions
• CBare-metal GPU access
• DObject storage replication

In RAG retrieval, increasing top-k from 3 to 20 typically

• AImproves recall but raises token cost and may hurt precision unless followed by reranking
• BAlways improves both precision and recall with no cost
• CReduces context length consumed
• DHas no effect on the prompt

Hybrid retrieval that combines BM25 with dense vector search is helpful because

• ABM25 captures exact keyword and rare-token matches that pure semantic vectors often miss
• BDense vectors are always worse than BM25
• CBM25 is faster than tokenization
• DIt removes the need for embeddings

A cross-encoder reranker such as bge-reranker is placed in a RAG pipeline

• AAfter initial retrieval to re-score and reorder the candidate chunks before sending the t…
• BBefore chunking the source documents
• CAs the embedding model itself
• DInside the LLM tokenizer

Query rewriting (a.k.a. query transformation) in RAG is useful when

• AThe user query is short, ambiguous or multi-part and needs to be expanded or split before…
• BThe vector database is full
• CYou want to disable the embedding model
• DThe user is typing in English only