cortexist/gemma4-decision-head

Gemma 4 classroom decision heads

Small CPU classifiers that map a Gemma 4 final-prompt hidden vector to speak, defer or wait. These are auxiliary heads, not language models or voice weights. They accompany the GGUF-based little-gemma v1.0 engine and little-gemma-tools v1.0 classroom. The classifier training revision is v2.1 (2026-10-05); it is independent of the software's v1.0 version. This package replaces gemma4-decision-head_v0.9.

Files

File	Use
gemma4-E4B-decision-head.gguf	E4B native C runtime head
gemma4-12B-decision-head.gguf	12B native C runtime head
e4b/head.npz, e4b/contract.json	E4B NumPy reference and feature contract
12b/head.npz, 12b/contract.json	12B NumPy reference and feature contract
SHA256SUMS	Checksums of release files

The GGUFs are byte-identical to the validated v2.1 exports, with plain public filenames. Their embedded contracts retain original training provenance. In particular, deployment_status records export-time status, not current runtime permission. E4B subsequently ran in the GGUF classroom; the supplied 12B revision is the latest offline-validated candidate, not a claim that every live 12B deployment used it.

Feature compatibility

Head	Hidden width	Layer	Required backbone GGUF SHA-256
E4B	2560	42	df0fd4ee07072c607c29a0a1cb4f98918426cca12f45a2776bdd6ee6d09a4de3
12B	3840	48	90fd44e29e0d7cffeb0fd00dc73cfdab9ed0b0e95306ecf7821ea634c940c370

Capture after output normalization at the final prompt token, before any generated-token forward, without answer-channel seeding. The input contract is trio_sound_event_v2_staged_history and task is final_prompt-action. Extraction used little-gemma revision c304f086a40c46ae4b67d5402738769dc12132a6; its serving support is included in v1.0. Matching dimensions alone is not enough: check backbone hash, layer, normalization, prompt format, and capture boundary. Other quantizations/backends, including EXL3, require separate validation or refitting. Do not bypass feature-contract checks to force an incompatible model.

Each head is a three-seed ensemble of 64-unit tanh MLPs with standardized, clipped inputs, mean logits and temperature softmax. Labels and numerical parameters are stored in the package; use them rather than hardcoding order. Training/dev/calibration support contains 5,751 authored examples, including 3,943 training rows across 37 training families. Probabilities are calibrated only on authored calibration families, not on unconstrained live conversations.

Use

hf download cortexist/gemma4-decision-head --local-dir decision-heads
cd decision-heads
sha256sum -c SHA256SUMS

Build the engine's decision_head shared library and decision-head CLI:

cmake -S little-gemma -B little-gemma/build -DCMAKE_BUILD_TYPE=Release
cmake --build little-gemma/build --target decision_head decision-head
little-gemma/build/decision-head inspect decision-heads/gemma4-E4B-decision-head.gguf

Configure the classroom's turn_head.backend as c, library as the absolute shared-library path, and each models entry as a head path plus its SHA-256. Use mode: control, text_decision_tags: true and prompt_lane: v2_sound_event with the compatible whisper_at_v2 acoustic adapter. The engine snapshots the prompt feature, runs the CPU head, then applies the decision to the same turn. The head itself does not synthesize audio, schedule speakers or detect overlap. See runtime configuration and classroom setup.

Evaluation and limitations

Authored panel	E4B	12B
Reserved invitation cases, three views	360 / 360	360 / 360
Previously exposed retention cases	1,212 / 1,212	1,211 / 1,212
Required-wait compact cases	40 / 40	40 / 40
Follow-up compact cases	30 / 30	30 / 30

Views are correlated renderings, not independent real-world trials. The 360-row panel contains 120 compact cases. The older retention set was exposed during development and is not a fresh holdout. The 12B retention error is a teacher-role case without sound metadata: a guest's own-name correction predicts Speak instead of Defer. Guest-role retention views passed. A previously exposed live E4B case was included in training and is not held-out evidence.

The original export check found identical C/NumPy choices on 6,111 vectors per model. These results establish narrow authored classroom behavior and export parity, not general conversational accuracy, interruption detection, live audio robustness or end-to-end latency. Sound metadata in the authored evaluation is synthetic; live ASR, identity, acoustics and prompt shifts can change outcomes. No private hidden vectors, transcripts, voice recordings or raw training corpus are distributed here.

Voices and licenses

Agent Voice provides the public guest voice separately. Users must choose two additional voices for teacher and student, prepare phoneme timing and split .enc.onnx/.dec.onnx graphs, and generate matching identity references. Chey's voice is private and closed-weight; it is not included or available for download.

The decision-head package is MIT licensed. Gemma backbones, voice models and external ASR/identity components retain their own terms; this license does not relicense those artifacts.