Beowulf Stroganoff: Building Economically Useless Chatbots
summary
A themed small LLM gets LoRA fine-tuning on a translated, thousand-plus-year-old Beowulf, a 17th-century cookbook, and The Story of Crisco, with ~1,000 generated question–answer pairs and five response categories. V1 follows a formula and returns recipes full of boiled blood and weird kennings; V2 adds slang, slogans, and curated prompts to reduce formulaic output and yields surprises like phenoxy comments and an unsolicited Mary Oliver line. The bot runs on a laptop GPU for training and lives on Hugging Face free tier (slow but usable).
AI-written summary · about
chapters (11)
- 0:00Introduction: What If We Made Language Models Bad?
- 1:33AI, Culture, and the Joy of Making Things
- 2:23Building Beowulf Stroganoff
- 3:23Fine-Tuning a Small Language Model with LoRA
- 4:20The First Results: Meatballs, Blood, and Phoxy Comets
- 5:10Improving the Training Data
- 6:37Designing Better Prompts and Responses
- 7:30Version 2: Weirdness with Intention
- 8:28Fine-Tuning vs. Programmatic Randomization
- 9:26Can AI Make Better Art?
- 10:04Protecting Creativity in the Age of AI
related talks
17mArchitecting LLM-Based Chatbots using free open-source LLMs
48mUnlocking LLM Potential: From Exploration to Integration
23mWriting Robot Plays NodeJS, Reddit
19mHow to get clean web data for chatbots and LLMs
01h 00mBuilding an Expert Q&A Bot with Open Source Tools and LLMs
01h 39mMastering LLMs: From Prompt Engineering to Agentic AI
01h 12mBuild Your Own AI Chatbot with OpenAI & LangChain
26mHow To Win Friends & Influence LLMs (with Prompt Engineering)