noFFENSE/gpt2-function-calling

GPT-2 function calling — before vs after

Side-by-side demo of GPT-2 355M implemented from scratch in PyTorch (no transformers) and fine-tuned for function calling on Glaive Function Calling v2 — against the untouched pretrained base model on the same prompt.

• —Code, tests, and training pipeline: <https://github.com/mron03/gpt2-function-calling>
• —Full write-up with animations: <https://mron03.github.io/gpt2-function-calling/>
• —Weights: <https://huggingface.co/noFFENSE/gpt2-355M-function-calling>

Runs on the free CPU tier with the hand-rolled KV cache — generation streams at a few tokens per second, one hand-written attention layer at a time.