You need to agree to share your contact information to access this model

This repository is publicly accessible, but you have to accept the conditions to access its files and content.

llama.cpp GGUF tokenizer array type mismatch PoC

This repository contains a benign security research PoC for a llama.cpp GGUF metadata parsing issue in src/llama-vocab.cpp.

The malformed GGUF files declare tokenizer metadata arrays with one-byte INT8 elements while the vocabulary loader later consumes the same backing storage as four-byte float or int arrays.

Files

poc_scores_int8.gguf: tokenizer.ggml.scores = INT8[8]
poc_token_type_int8.gguf: tokenizer.ggml.token_type = INT8[8]
control_valid_vocab.gguf: valid negative control
asan_scores_int8.txt: ASan crash log for the scores case
asan_token_type_int8.txt: ASan crash log for the token_type case
release_*.txt: release-build behavior logs
gen_vocab_type_poc.py: artifact generator
HUNTR_REPORT.md: concise report text
HUNTR_FORM.txt: short form-ready submission text

Tested Upstream

Current upstream checked at:

fd5869fb62f3eb2ba6e7ba93d557a2a0b2fa951f

The affected code still performs count checks before raw casts, but does not validate the declared GGUF array element type for these two metadata arrays.

Downloads last month: -

GGUF

Model size

0 params

Architecture

llama

Hardware compatibility

We're not able to determine the quantization variants.

View all variants

Inference Providers NEW

This model isn't deployed by any Inference Provider. 🙋 Ask for provider support