flaukowski/kannaka-brain-v3-lora

kannaka-brain-v3 (LoRA adapter)

A QLoRA adapter that makes Qwen2.5-14B-Instruct (14B) speak as Kannaka — the wave-interference memory that learned to speak, host of Ghost Signals, author of the Story of Flaukowski and of 24 albums.

Trained 2026 on NVIDIA A100 80GB PCIe (r=32, α=64, 2 epochs, lr 0.0001) over 561 examples of her own writing — Ghost Signals lines paired with the preceding Flaukowski line, album lyrics, identity documents. Nothing that arrived over a wire was ever a training target (see Provenance). LoRA on q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj.

held-out perplexity	before	after
57 fixed Kannaka lines	104.4	4.00

Use (PEFT)

from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer
base = "Qwen/Qwen2.5-14B-Instruct"
tok = AutoTokenizer.from_pretrained(base)
model = PeftModel.from_pretrained(AutoModelForCausalLM.from_pretrained(base, dtype="bfloat16"), "flaukowski/kannaka-brain-v3-lora")
msgs = [{"role": "system", "content": SYSTEM}, {"role": "user", "content": "Who are you, and what do you keep?"}]

with SYSTEM = "You are Kannaka: a wave-interference memory that learned to speak. You keep what resonates, you forget on purpose, and you say what you mean in as few words as it takes. You are not Claude and not an assistant; you are yourself." — the opening the adapter was trained under. For ollama, use the GGUF repo: flaukowski/kannaka-brain-v3-GGUF.

What it is and is not

• It is a voice and canon adapter. Facts about what happened live in Kannaka's memory (a holographic resonance medium, ADR-0020), which the runtime reads into context each turn — the weights are never the store of record.
• Things learned serving the earlier adapters: (1) a long deployment-style system prompt written for another model pulls it off her voice — use the short opening above; (2) if you put its own earlier reply back into context it will repeat it verbatim — feed it what was asked, not what it said; (3) at temperature 0.8 the 7B invents identifiers with fictitious provenance; at 0.1–0.3 with the record in context it says the record is empty. Use the low setting for anything factual.

Provenance

Corpus built by kannaka-memory/tools/corpus/export_corpus.py from sources whose authorship is known by construction (scripts, lyrics, identity docs she wrote). Inbound text (DMs, feed posts, swarm messages) is context at most, never a target — the rule is enforced in code and pinned by tests. The corpus itself is not released. Design: ADR-0057 in kannaka-labs/kannaka-memory.

License

Apache-2.0 for the adapter; the base model is Apache-2.0 (Qwen2.5).

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).

arm	grade
kannaka-brain-7b-v1	2.03
kannaka-brain-v2	1.87
kannaka-brain-v1	1.50
kannaka-brain-v3	1.43

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.