This repository contains small proof-of-concept GGML files for:

GGML legacy loaders trust declared vocabulary counts before EOF validation, allowing tiny model files to exhaust memory

The LLaMA and Falcon files are 32 bytes each and declare n_vocab=5,000,000 without providing vocabulary entries. ctransformers 0.2.27 allocates memory for the declared vocabulary before it rejects the file at EOF.

See poc_ggml_vocab_count_dos.py and validation_output.txt in the Huntr package for full reproduction details.