Lessons
view all ⭢
lesson
Evaluating Preference AlignmentAI BootcampApply ensemble voting to eliminate hallucinations, compare DPO vs RLHF vs PPO pipelines (cost, control, complexity).
lesson
Preference-Based FinetuningAI BootcampUnderstand DPO, PPO, RLHF, GRPO; generate math-focused DPO datasets using numeric correctness.
lesson
Tool-Augmented FinetuningAI BootcampImplement two-stage finetuning (CoT → CoT+Tool), evaluate reasoning accuracy, and train quantized LoRA models.
lesson
Math Reasoning with SymPyAI BootcampUse SymPy for symbolic reasoning, fine-tune with Chain-of-Thought (CoT) data blending natural language and Python.
lesson
Deploying Finetuned CLIP ModelsAI BootcampEvaluate accuracy with cosine similarity, deploy in insurance workflows with LLaMA for reasoning.
lesson
CLIP Finetuning for InsuranceAI BootcampFine-tune CLIP for car damage classification, use Google Custom Search API for collecting datasets, and apply LoRA with Optuna for optimization.
lesson
RAG Evaluation and OptimizationAI BootcampEvaluate RAG outputs (recall, MRR, qualitative), optimize retriever+generator coordination for enterprise use cases.
lesson
Advanced RAG SystemsAI BootcampAnalyze production-grade RAG case studies (Relari, Evidently), understand bottlenecks (chunking, reranking), and compare embedding models.
lesson
Monkeywrenching into LLaMAAI BootcampEdit layers in a pre-trained model, understand a state-of-the-art model and monkey patching cutom layers in State of the art model.
lesson
Debugging & Testing TransformersAI BootcampUse sanity checks and test loss to debug, observe transformer behavior on structured prompts/simple sequences.
lesson
Building a Full TransformerAI BootcampConnect embeddings, attention, feedforward, and normalization; implement skip connections and positional encodings manually.
lesson
Mini-Project: Generating Synthetic Data & Finetuning for EvaluationAI BootcampMini Projects: Generating Different Synthetic Data & Fine-Tuning for Evaluation-Centric Applications w.r.t Building synthetic data for non-reasoning, Building synthetic data for reasoning, and Building synthetic data for diversity.
