Released as part of the NOESIS Professional Multilingual Dubbing Automation Platform (framework: DHCF-FNO — Deterministic Hybrid Control Framework for Frozen Neural Operators).
✅ WIRED via Path A G4→G3 aligner v1 (2026-06-05). The plug-in
incompatibility described below (R-SCENEMA-GEMMA4-PLUG-IN-INCOMPATIBLE)
is now bridged by a trained non-linear aligner — see
scenema_aligner_v1/ in this bundle. The aligner maps Gemma-4's
output hidden states onto the Gemma-3 distribution that Scenema-DiT
expects. Pipeline: Gemma-4 forward → aligner(hs_g4) → Scenema-DiT.
Quality (Path A v1, step 12000, full clean run RC=0):
mean per-lang val_cos = 0.908; 12 of 13 languages ~0.92
(es,pt,de,it,ru,fr,en,sw,hi,zh ≈ 0.92-0.93; ja ≈ 0.910; ar ≈ 0.899);
ko (Korean) = 0.728 — weak outlier, gate_passed=false is driven
by Korean ONLY. Accepted as-is by operator 2026-06-05 (Korean fix
deferred to a future korean-weighted retrain; the aligner is usable for
the other 12 languages now).
The aligner is a NON-LINEAR module (LayerNorm → Linear-GELU-Linear +
residual → per-layer scale/bias). It is applied to Gemma-4 OUTPUT hidden
states at inference and CANNOT be weight-merged into Gemma-4's
model.safetensors — folding a GELU/LayerNorm into transformer weights
is mathematically impossible. The "merge" is a paired runtime: the
Gemma-4 weights + the aligner sidecar, wired in sequence.
NOESIS provenance
Bundle : NOESIS-Gemma4-12B-it-Qat-Q4_0-Unquantized-BF16
Parent bundle : NOESIS-3.5B-A0.5B-DUBBING-FILM /
LTX2.3-5B-Scenema-Audio-DiT-BF16
Upstream : google/gemma-4-12b-it (Google LLC, Gemma Terms of Use)
NOESIS variant : Verbatim upstream BF16 weights (NOT modified) + PAIRED
Path A G4→G3 aligner v1 (scenema_aligner_v1/).
NOESIS role : Wired Scenema text-encoder (G4 path). Path 1 of the
three GEMMA4_INTEGRATION_HANDOFF_2026-06-04 paths —
G4→G3 hidden-state aligner — was TRAINED and accepted:
1. ✅ G4→G3 aligner v1 (DONE 2026-06-05, mean 0.908,
12/13 langs ~0.92, ko=0.728 weak; trained on AWS
L40S, full 12000 steps RC=0)
2. Retrain audio_aggregate_embed ($30-80) — not done
3. Route 2 v4 with G4 teacher ($20-50) — not done
The aligner resolves the plain-swap incompatibility for
12 of 13 languages. Korean (ko=0.728) deferred.
Downloaded : 2026-06-04
Last updated : 2026-08-29 (Path A G4→G3 aligner v1 trained + paired;
status: held-for-experiments → WIRED)
================================================================================
⚠️ License gate (Gemma Terms of Use)
Use is conditioned on accepting the Gemma Terms and complying with
the Prohibited Use Policy.
Commercial use IS permitted under the Gemma Terms (no revenue cap)
but Google's restrictions on use, redistribution, and derivative
works still apply.
Redistribution must include a copy of the Gemma Terms and provide
prominent attribution to Google as the upstream Licensor.
Any NOESIS-side fine-tune or adapter built on these weights becomes
a Derivative Work subject to the same Terms.
This bundle is the unmodified upstream BF16 weights as published by
Google. NOESIS has not modified the weights — see
R-SCENEMA-GEMMA4-PLUG-IN-INCOMPATIBLE-WITHOUT-DIT-RETRAIN and the
Route 2 experiment handoff for the planned downstream LoRA / adapter
training paths.
================================================================================
Why this is held (not yet usable in production)
Per CLAUDE.md (UPD 2026-06-04 evening, R-SCENEMA-ROUTE2-ADAPTER-V3-RU-FAILS-LOCAL
Gemma 4 integration handoff):
Property
Gemma 3 12B (current teacher)
Gemma 4 12B (this bundle)
Shape: [49 × 3840 → 188160]
✅ matches
✅ matches
Per-layer hidden-state stacked-norm
~10,267,738
~23,464 (×437 smaller)
Per-channel hidden-state std
up to ~13
up to ~163,295 (×12,500 larger)
Plain plug-in into Scenema DiT
✅ works (R-SCENEMA-TEXT-ENCODER-PAIRED)
❌ fails — Russian phonemes only, no content words
norm_and_concat_per_token_rms alignment
n/a
partial — amplitude lost, alignment works on phonemes only
The shape compatibility is misleading: Scenema-DiT was trained on
Gemma 3 hidden-state amplitude distributions. Gemma 4 substitution
produces audio that contains real Russian phonemes but not the words
of the input prompt. The fix requires either:
G4→G3 hidden-state alignment adapter (cheapest, recommended):
train a small Linear / affine adapter that maps Gemma 4 last hidden
layer onto the Gemma 3 distribution that Scenema-DiT expects.
$5-15, A100 80GB, ~3-5h.
Retrain audio_aggregate_embed: regenerate the frozen
projection in scenema-audio-pipeline.safetensors to expect Gemma
4 distributions natively. $30-80.
Route 2 v4 with Gemma 4 teacher: re-run the Route 2 cross-lingual
KD with Gemma 4 as teacher (instead of Gemma 3). $20-50.
All three paths are CLOUD-SIDE experiments — none fit on the RTX 3060
6 GB local node.
1# from inside scenema_aligner_v1/2from apply_aligner import load_aligner
3aligner = load_aligner("aligner_g4_to_g3_v1.pt", device="cuda")45# hs_g4 = Gemma-4-12B last-hidden-state stack [B, T, 3840]6hs_aligned = aligner(hs_g4)# mapped onto Gemma-3 distribution7# feed hs_aligned into Scenema-DiT cross-attention as the text condition
The aligner is applied at INFERENCE between Gemma-4's forward pass and the
Scenema-DiT — it is NOT merged into model.safetensors (non-linear, cannot be
folded into transformer weights). Korean (ko) output is weak (0.728); the other
12 languages are production-usable (~0.92).
================================================================================
NOESIS Scenema Family
R-SCENEMA-GEMMA4-ALIGNER-V1-WIRED-2026-06-05
(sealed 2026-06-05, SUPERSEDES the "plug-in incompatible" status)
The G4→G3 hidden-state aligner (Path 1) was TRAINED and accepted:
scenema_aligner_v1/aligner_g4_to_g3_v1.pt (253 MB, non-linear MLP +
per-layer scale/bias). Full 12000-step run on AWS L40S, RC=0.
Results: mean per-lang val_cos 0.908; 12/13 langs ~0.92;
ko (Korean) = 0.728 (weak — gate_passed=false driven by Korean only).
Operator accepted as-is 2026-06-05; Korean fix deferred to a future
korean-weighted retrain. The aligner is applied at inference
(Gemma-4 forward → aligner → Scenema-DiT) and is NOT weight-merged
into model.safetensors (non-linear, cannot fold into transformer
weights). This resolves the previous plain-swap incompatibility for
12 of 13 languages.
R-SCENEMA-GEMMA4-PLUG-IN-INCOMPATIBLE-WITHOUT-DIT-RETRAIN
(sealed 2026-06-04 — HISTORICAL, now bridged by aligner v1 above)
Gemma 4 12B CANNOT PLAIN-swap into Scenema-DiT (shape matches but
hidden-state distribution differs ×437 stacked-norm / ×12,500 std).
Bridged by R-SCENEMA-GEMMA4-ALIGNER-V1-WIRED-2026-06-05 (aligner v1).
R-GEMMA-TERMS-PRESERVE
Upstream Gemma Terms of Use preserved verbatim in LICENSE.
Bound by Google's Gemma Terms + Prohibited Use Policy.
R-HELD-FOR-EXPERIMENTS
Held on disk as the substrate for cloud-side Route 2 G4→G3
adapter experiments. Not yet a NOESIS LoRA fine-tune. NOESIS-side
LoRA / adapter training will happen on cloud (out of scope for
RTX 3060 6 GB local node).
R-DUBBING-FILM-SCOPE (sealed 2026-04-29)
NOESIS = professional audio dubbing. This bundle is research /
experiment substrate for the Scenema text-encoder upgrade path.
R-NEVER-DELETE-WITHOUT-EXPLICIT-CONSENT (sealed 2026-05-21)
MUST NOT be deleted without explicit operator instruction
"удали " / "delete ". Held under R-NEVER-DELETE because
cloud Route 2 adapter experiments still pending — losing this
substrate forfeits the G4→G3 alignment retry path.
