Small Language Models on Device: Phi, Gemma, Llama 3.2 in Production

End-of-lesson check

15 questions · take it digitally for instant feedback at tendril.neural-forge.io/learn/quiz/end-model-families-small-language-models-on-device-creators

1. A mobile app developer needs to classify user feedback into positive, negative, or neutral categories entirely on the user's device to ensure user data never leaves the phone. Which solution best fits this need?

   1. A human review pipeline for all feedback
   2. A cloud-based machine learning service with encryption
   3. A small language model running directly on the device
   4. A frontier model accessed via API call for maximum accuracy

2. What is the primary reason production systems often combine small language models with frontier model APIs?

   1. It creates a competitive advantage over using only one model type
   2. Government regulations require using both model types together
   3. Different tasks have different complexity requirements, and each model type excels at different things
   4. Small models are more accurate than frontier models for all tasks

3. A company must process sensitive financial documents that cannot leave their secure data center due to strict compliance regulations. Which characteristic of on-device SLMs addresses this requirement?

   1. Their ability to connect to multiple APIs simultaneously
   2. Their ability to generate creative marketing content
   3. Their ability to scale horizontally across data centers
   4. Their ability to run completely offline without external connections

4. You are building an autocomplete feature that suggests the next word as a user types in a document editor. Speed is critical, and users expect instant results. Why would an on-device SLM be preferable to calling a frontier model API?

   1. On-device SLMs have unlimited context windows
   2. On-device SLMs eliminate network round-trip latency entirely
   3. Frontier models cannot generate single-word predictions
   4. Frontier models require paid subscriptions

5. Your application needs to summarize a 50-page legal contract and identify all instances of specific clause types. What should you consider before deciding between an SLM and a frontier model?

   1. The color of the document
   2. Whether the contract is written in English or another language
   3. The model's context window limitations and the complexity of the extraction task
   4. Whether the contract contains signatures

6. A startup is building an AI feature that needs to answer open-ended questions about philosophy and ethics with nuanced, thoughtful responses. Which model approach would likely fail to meet this need?

   1. A small language model running on-device
   2. A hybrid system with SLM for preprocessing
   3. Any cloud-based model service
   4. A frontier model accessed via API

7. Which three model families were specifically named in the lesson as examples of small language models suitable for on-device deployment?

   1. Claude, GPT, and Gemini
   2. Mistral, Falcon, and Stable LM
   3. T5, BERT, and RoBERTa
   4. Phi, Gemma, and Llama 3.2

8. A company deploys an SLM to their mobile app in January. In March, a major new technology breakthrough is announced that changes industry standards. What limitation of SLMs will prevent their model from incorporating this new information?

   1. SLMs have built-in knowledge cutoffs that update automatically
   2. SLMs are connected to live internet feeds
   3. SLMs automatically update their knowledge base weekly
   4. SLMs cannot learn from new data without retraining

9. An e-commerce platform processes millions of product review classifications daily to categorize feedback for internal dashboards. Cost control is critical. Why might an SLM be the better choice over a frontier API for this task?

   1. SLMs are more accurate than frontier models for classification
   2. Classification tasks require internet connectivity
   3. Frontier models cannot perform classification tasks
   4. SLMs can run locally after one-time deployment, eliminating per-request API costs

10. What does the lesson recommend doing before deciding to route a specific task to an SLM in a production system?

    1. Check the model's parameter count
    2. Deploy it to all users immediately and monitor complaints
    3. Ask users which model they prefer
    4. Measure quality on your evaluation set using both SLM and frontier approaches

11. A healthcare application needs to extract patient symptom codes from doctor notes while ensuring HIPAA compliance. The notes must never be transmitted over networks. Which architectural choice best satisfies these requirements?

    1. Use human transcriptionists for all notes
    2. Run an SLM on secure servers within the healthcare facility's network
    3. Send notes to a frontier API with HIPAA-compliant security
    4. Process notes using a cloud-based NLP service

12. What is the typical context window limitation for most small language models mentioned in the lesson?

    1. Up to 128K tokens
    2. Up to 1K tokens
    3. Unlimited tokens
    4. 8-32K tokens

13. In a hybrid production system architecture, which task would typically be routed to an SLM rather than a frontier model?

    1. Classifying an incoming support ticket into predefined categories
    2. Generating a creative short story with complex plot twists
    3. Writing a novel with nuanced character development
    4. Answering a philosophical debate about consciousness

14. A developer is building an offline-first mobile keyboard app that suggests emoji and text completions. The app must work in airplane mode. Which model characteristic makes this possible?

    1. Only frontier models support text completion
    2. SLMs can run entirely on-device without network connectivity
    3. SLMs require constant cloud synchronization
    4. Frontier models have smaller file sizes than SLMs

15. What architectural component should be planned from day one in a system that uses both SLMs and frontier models?

    1. The routing layer that decides which model handles each request
    2. The marketing budget allocation
    3. The company logo design
    4. The user interface color scheme