You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
An experiment-driven laboratory for understanding LLM sampling strategies. Explore how Temperature, Top-k, and Top-p influence token probabilities and generation diversity through reproducible simulations, visualizations, and real Groq API experiments.
The LLM Defense Framework enhances large language model security through post-processing defenses and statistical guarantees based on one-class SVM. It combines advanced sampling methods with adaptive policy updates and comprehensive evaluation metrics, providing researchers and practitioners with tools to build more secure AI systems.
Orchestrator-worker multi-agent pattern: one planner, three independent workers with identical prompts but different sampling configs (temperature/top_p). Isolates what sampling alone changes in model behavior — conservative vs balanced vs creative brainstorming.
A custom autoregressive text generation engine with manually implemented Greedy, Temperature, Top-K, and Top-P (Nucleus) sampling strategies, plus a from-scratch KV-Cache implementation benchmarked for inference speedup.
Medical Q&A chatbot: BioBERT encoder + Transformer decoder built from scratch in PyTorch. Beam search and nucleus sampling. Research and education only.