LLM_from_scratch (github.com)

LLM_from_scratch is a comprehensive, hands‑on walkthrough and codebase that teaches how to build, train, fine‑tune and align a causal large language model end‑to‑end using PyTorch. It starts with environment setup (conda, CUDA/Mac, mixed precision and profiling), implements transformers from first principles (positional embeddings, manual self‑attention → multi‑head attention, MLPs with GELU, residuals and LayerNorm), and stitches those into full models. The guide then covers dataset/tokenization choices (byte‑level → BPE), batching/label shifting for next‑token prediction, a bare‑metal training loop (no Trainer API), sampling (temperature, top‑k/top‑p), evaluation, and practical infra (gradient accumulation, mixed precision, LR schedules, checkpointing, TensorBoard/W&B). Beyond the basics it dives into research‑grade extensions and deployment techniques: architectural variants (RMSNorm, RoPE rotary embeddings, SwiGLU), inference speedups (KV cache, rolling/ sliding‑window attention), MoE layers for scaling, and hybrid dense/MoE designs. For alignment it covers instruction formatting, preference datasets, reward model training, and RLHF via PPO (KL‑penalized objective, value head) plus an alternative GRPO group‑relative baseline (per‑prompt reward normalization, policy‑only clipped loss with explicit KL penalty). The package is significant because it pairs pedagogical clarity with runnable code, showing implementation details and practical tradeoffs that researchers and engineers need to reproduce, debug, and iterate on modern LLMs.