Show HN: Create-LLM – Train your own LLM in 60 seconds (github.com)

Create-LLM is a one-command scaffolding tool that generates a full, production-ready PyTorch pipeline for training custom language models—think create-next-app but for LLMs. Using npx @theanikrtgiri/create-llm <project>, you get tokenizer training (BPE/WordPiece/Unigram), data preprocessing, trainer with callbacks, checkpoint management, TensorBoard/live dashboard, interactive chat, evaluation/generation scripts, and deployment helpers (Hugging Face, Replicate). Templates range from NANO (≈1M params, CPU-friendly, minutes) through TINY (≈6M), SMALL (≈100M, single RTX3060-scale), to BASE (≈1B, multi-GPU/A100), with recommended hardware, data-size and time estimates included. The generated project is configured via llm.config.js, supports plugins (WandB, SynthexAI), auto-detects tokenizer vocab sizes, warns about model/data mismatches and overfitting, and provides resume, compare, and dashboard capabilities. This lowers the barrier to entry for learning, rapid prototyping and small-scale production models by packaging best practices (checkpointing, eval, tokenizer training) into a reproducible workflow. For researchers and practitioners it speeds iteration and reproducibility; for educators it’s a practical teaching tool. Caveats: current emphasis is GPT-style architectures, limited distributed/quantization support, and real performance depends on data quality and hardware—small templates can overfit easily and BASE-scale models still require substantial compute and data. Overall, Create-LLM is a pragmatic, opinionated starter kit that accelerates experimentation while highlighting the usual trade-offs of data, compute, and model size.