Set the model name and min_p in the GGUF metadata (weights unchanged)
A metadata-only update to all five GGUF files. The weights are byte-for-byte identical to the current files: only the header's key-value table changes.
What changes:
general.name becomes Ternary-Bonsai-2-27B in every file. It previously read Hf.
general.basename and general.version are removed from the three text files, where they read folded and v5. They were picked up from a build directory name and don't describe the model.
general.sampling.min_p = 0.0 is added to the three text files (PQ2_0, PTQ1_0, F16), matching Qwen's recommended thinking-mode settings for the base model. The files already carried top_k, top_p and temperature, but without min_p llama.cpp fell back to its default of 0.05.
How it was checked:
Each patched file was built from the current file, whose whole-file SHA-256 matched the repository's recorded hash, and the tensor-data section's SHA-256 is identical before and after.
A diff of the key-value tables shows exactly the changes above and nothing else. Tensor counts are unchanged.
With the llama.cpp release build, PQ2_0 and PTQ1_0 give identical perplexity before and after on the same text, and the sampler reports min_p = 0.000 with the patched files instead of 0.050.
Each file grows or shrinks by a few dozen bytes, and every file's hash changes. For reference, the previous SHA-256 values are: PQ2_0 3907dc1658db1f78a9826bf8d5bcb8dc65db0d466388937af57f2294fae62ec1, PTQ1_0 53107f530aa52eb00912263ab1ee29bd199261c87cd7b4ad4ca1318c1fe33ee3, F16 f6f3b2c9b41956c34b379ec7c301dc936bc38d79b3c24c83388dd7d76000c180, mmproj-BF16 e287342d92332fa3577ed1d42e921dac9370c08da58ba9337fa450f6cc76cfd7, mmproj-Q8_0 6807ede61d570bb86ba34b756a0fa109edc33668604de867c6ea6d8f1d631903.
