Syllabus Map

• Study map: Syllabus Study Map

Overview

• Pre-trained language models can be open-source or API-based.
• They are trained on large corpora and adapted via prompting, fine-tuning, or retrieval augmentation.

Model Categories

Base (Pretrained) Models

• Strong token prediction capabilities.
• Usually require task-specific tuning for best performance.

Instruction-Tuned / Chat Models

• Further tuned on instruction-response data.
• Better at following natural language tasks out of the box.

Reasoning-Optimized Variants

• Often tuned for multi-step problem solving.
• Trade some latency/cost for stronger complex-task behavior.

Open-Source Families

• LLaMA family.
• Mistral family.
• Qwen family.
• Typical deployment options:
• full precision for maximum quality,
• quantized (for example 4-bit/8-bit) for lower memory and faster local inference.

API-Based Models

• GPT-style APIs.
• Claude-style APIs.
• Managed APIs reduce infra burden and provide rapid model upgrades.
• Tradeoff is less control over model internals and serving environment.

Selection Criteria

Capability Fit

• Evaluate on your real tasks: summarization, extraction, coding, classification, etc.
• Match model size/context length to task difficulty and document length.

Cost and Latency

• Measure end-to-end latency at target throughput.
• Include token pricing, retries, and context-window overhead.

Safety and Governance

• Check prompt injection resistance, refusal behavior, and data handling requirements.
• Add policy layers and output filters for production.

Deployment Constraints

• API model: fastest integration.
• Open-source model: more control, local/privacy-friendly, but higher ops complexity.

Practical Adaptation Patterns

Prompting Baseline

• Start with strong prompt templates and few-shot examples.
• Establish baseline before heavier adaptation.

RAG (Retrieval-Augmented Generation)

• Retrieve relevant documents and append to context.
• Reduces hallucinations in domain-specific QA.

Fine-Tuning

• Use supervised fine-tuning for style/task alignment.
• Prefer parameter-efficient methods (LoRA/QLoRA) when compute is limited.

Practical Notes

Model Selection

• Choose based on latency, cost, and data constraints.

Evaluation Discipline

• Always evaluate on a held-out task set with failure-case slices.
• Track regression across model/version changes.