Build these to solidify each phase
Phase 1 — Foundations
Implement a multi-layer perceptron in pure NumPy. Train on MNIST. Understand backprop at the math level.
Phase 2 — NLP Core
Fine-tune a BERT model on IMDB reviews. Full pipeline: data loading, tokenization, training, evaluation with HuggingFace.
Phase 2–3 — Transformers
Implement a GPT model in PyTorch following Karpathy's nanoGPT. Train on Shakespeare or OpenWebText.
Phase 5 — Applications
Build a retrieval-augmented chatbot. Chunk documents, embed with sentence-transformers, store in a vector DB, and query with an LLM.
Phase 4 — Fine-Tuning
Fine-tune a 7B+ model on custom instruction data using QLoRA. Evaluate with lm-eval-harness. Serve with vLLM.
Phase 5 — Agents
Build an agent that searches the web, writes code, calls APIs, and reasons over results. Implement ReAct or function-calling patterns.
Phase 3 — Pre-training
Curate data from CommonCrawl/RedPajama, build a BPE tokenizer, train a 100M–1B param model. Full distributed training pipeline.
Phase 4 — Alignment
Collect preference data, implement DPO/RLHF training with TRL, compare alignment quality vs SFT baseline on MT-Bench.
Phase 5 — Applications
Build a semantic search system: embed a corpus with sentence-transformers, index with FAISS/Qdrant, add LLM-powered answer synthesis.
Phase 6 — Research
Build an app using vision-language models (LLaVA, Qwen-VL). Support image Q&A, OCR, and visual reasoning.
Phase 2 — Getting Started
Run open-source LLMs locally with Ollama. Build a terminal chatbot with streaming, system prompts, and conversation memory.
Phase 5 — Deployment
Quantize a model (GPTQ/AWQ), serve with vLLM behind an API, add monitoring with Prometheus + Grafana. Load-test with Locust.
Phase 1–2 — Foundations
Implement Byte-Pair Encoding in Python. Understand how text becomes tokens — the first step in every LLM pipeline.
Phase 2–3 — NLP + LLMs
Build an extractive + abstractive summarization system. Compare BART, T5, and Pegasus models on CNN/DailyMail.
Phase 3–4 — Evaluation
Set up lm-evaluation-harness. Benchmark models on MMLU, HellaSwag, TruthfulQA. Compare base vs fine-tuned models.
Phase 4 — Optimization
Distill a large model into a smaller one. Implement logit matching and feature alignment. Measure quality vs speed trade-offs.
Phase 2–5 — Prompting
Build a web UI to test prompt techniques: zero-shot, few-shot, CoT, ReAct, self-consistency. Side-by-side model comparisons.