Building The Language Model Nobody Asked For
a day ago
- LLMs like ChatGPT are trained to be intelligent and polite, but lack humor and spontaneity.
- The author aims to create a more interesting LLM by fine-tuning an open-weights model with nontraditional data.
- After comparing model families (Llama, Qwen, Mistral), Mistral Small 3 24B is chosen for its open access and trainability on consumer GPUs.
- Mistral Small 3 avoids synthetic data and RL training, producing more human-like text.
- Training data is sourced from websites that contrast with corporate speak, such as humorous or chaotic content.
- The resulting model is described as both smarter and funnier than ChatGPT, with weights available for download.