Find things worth doing in local AI
Differential chat-template auditing for the local model supply chain
7469% evidence
Build the missing conformance gate for chat templates in local model artifacts: a tool that extracts `tokenizer.chat_template` from a GGUF or HF repo, normalises and diffs it against the canonical upstream template for that model family, renders both across a probe corpus (multi-turn, system prompts, tool definitions, special tokens) under each runtime's own Jinja implementation, and diffs the resulting token streams. Divergences are classified — benign reformatting, BOS/EOS duplication, tool-call notation drift, or the high-signal case of template logic that branches on user-supplied content, which is the signature of an inference-time backdoor. Ships as a CLI plus CI action for quant publishers, with a hub-wide scan as the launch artifact. Templates are small text files and rendering needs no GPU, so ecosystem-scale coverage is cheap.
read the full report →
