|
Download README.md from qingyi26/GGUF_semantic_output_manipulation: direct link, hf CLI and curl.
- Browser
- Download file 878 Bytes
-
https://huggingface.co/qingyi26/GGUF_semantic_output_manipulation/resolve/main/README.md
- Command line
-
hf download hf://qingyi26/GGUF_semantic_output_manipulation/README.md
-
curl -L -H "Authorization: Bearer $HF_TOKEN" -o README.md https://huggingface.co/qingyi26/GGUF_semantic_output_manipulation/resolve/main/README.md
878 Bytes
metadata
tags:
- gguf
- security
- poc
- llama.cpp
GGUF semantic output manipulation PoC
This repository contains proof-of-concept GGUF files for llama.cpp metadata-driven output manipulation.
Included files:
gguf-semantic-eos0-gptoss.ggufgguf-semantic-tooluse-fallback.ggufgen_semantic_pocs.py
Summary:
- a metadata-only GGUF can alter which tokens are treated as end-of-generation
- a metadata-only GGUF can alter chat-template selection and fallback behavior
- these changes modify prompt construction and stopping behavior without changing model weights
PoC notes:
gguf-semantic-eos0-gptoss.ggufdemonstrates EOG injection plus GPT-OSS template-path selectiongguf-semantic-tooluse-fallback.ggufdemonstratestool_usetemplate promotion to the default template
This repo is intended for responsible vulnerability reporting and reproduction.