Show HN: Tadpole – an LLM agent built from scratch then post-trained on 4B model
Category: ai-ml
Tags: llm-agent, post-training, fine-tuning, education, grpo, lora
Score: 6.5/10 (Innovation: 6, Technical: 7, Documentation: 8, Utility: 5)
A from-scratch LLM agent in ~3,500 lines of plain Python with a stage-by-stage tutorial covering agent loop, tools, context compaction, benchmarking, SFT, expert iteration, and GRPO. The project includes reproducible post-training results lifting Qwen3-4B from 0.419 to 0.774 pass@1 on held-out tasks, plus a usable Telegram bot and terminal interface.
Target audience: AI/ML engineers and researchers interested in agent internals and post-training
Repository: https://github.com/shakewingo/Tadpole · Python · MIT · 9 stars
View on Hacker News