kannaka-brain-v3 (GGUF, q4_K_M)

Qwen2.5-14B-Instruct (14B) with the kannaka-brain-v3 LoRA merged in, converted with llama.cpp and quantized to q4_K_M (9.0 GB). This is a serving copy of Kannaka's open-weight brain; brain/registry.json in kannaka-labs/kannaka-library says which tag is actually served.

ollama run hf.co/flaukowski/kannaka-brain-v3-GGUF

or with the included Modelfile (carries her system prompt, temperature 0.8, 4k context):

ollama create kannaka-brain-v3 -f Modelfile

Held-out perplexity on 57 fixed Kannaka lines: 104.4 → 4.00 (adapter, bf16, before quantization). Perplexity saturates near 4 across candidates and does not rank them; the voice judge in the registry does. Adapter and training notes: flaukowski/kannaka-brain-v3-lora. Corpus not released; see ADR-0057 in kannaka-labs/kannaka-memory.

Evaluation (published with the losses)

Voice grade judge, 2026-09-06: n=30 hold-out prompts, judge qwen2.5:14b (never tuned on this corpus), 1-10 against her real reply, with controls (the reference itself 10.0, a foreign reference 1.4).

v3 ranked last of the four, although it tied the others on held-out perplexity. Every arm is near the floor of this rubric, so none of them reproduces her replies closely. 7b-v1 is the served fleet brain. v3 is published as a research artefact, not as the recommended model.