Project P3

Hyperparameter Playground

Concept: LLM Sampling & Logits Probability

Problem Statement

Tuning LLMs is often treated as guess-and-check. Developers do not understand how Temperature or Top-p parameters affect output predictability.

What Concept It Teaches

It teaches Softmax logit curves scaling, Top-p/Top-k selection bounds, and deterministic response bounds.

Why This Matters

Fine-tuning parameters changes output quality, ensuring stable JSON responses in schemas and high creativity in copywriting.

System Architecture

Logs visualizer showing vocabulary probability bars changing dynamically as sliders scale parameters.

Raw Logits ArrayTemperature Scale FunctionSoftmax Probability ConverterTop-p/Top-k Filter GateToken Output Pick

Execution Data Flow

  • 1. Preset logits are loaded for a next-token choice.
  • 2. Sliders scale Temperature and Top-p limits.
  • 3. Probability bar charts redraw in real-time.
  • 4. UI shows the selected token vs candidates.

Tech Stack

ReactTypeScriptTailwind CSS

Implementation Plan

  • 1.1. Build Softmax math utilities.
  • 2.2. Render dynamic bar graphs of vocabulary sets.
  • 3.3. Implement Top-p selection bounds filters.

Technical Interview Defense

Defense Concept:

Derive Temperature scaling inside Softmax, explaining why lower temperatures yield predictable responses.

Source & Deployment Links

GitHub Repo:GitHub
Live Demo:Live Demo
Verification Audit
Repository Checked: Yes
Repository Exists: Yes
Live Demo Verified: Yes
Demo Exists: Yes