NOESIS / AMAImedia
Released as part of the NOESIS Professional Multilingual Dubbing Automation Platform (framework: DHCF-FNO — Deterministic Hybrid Control Framework for Frozen Neural Operators).
- Founder: Ilia Bolotnikov
- Organization: AMAImedia.com
- X (Twitter): @AMAImediacom
- LinkedIn: Ilia Bolotnikov
- Telegram: @djbionicl
- NOESIS version: v16.1
- Release date: 2026-08-26
================================================================================ NOESIS-Gemma3-1B-BF16 -- NOESIS Bundle README
NOESIS LoRA-merged Scenema student (1B class). Authoritative BF16 master for the lightweight Scenema-DiT audio cross-attention text encoder (recipe-validation companion / smoke-test student for the primary 4B Scenema student bundle).
NOESIS provenance
Bundle : NOESIS-Gemma3-1B-BF16
Parent bundle : NOESIS-3.5B-A0.5B-DUBBING-FILM /
LTX2.3-5B-Scenema-Audio-DiT-BF16
Folder lineage : Renamed 2026-06-04 (was
NOESIS-Gemma3-1B-PhaseE-Merged-BF16; renamed per
v4-archive folder-rationalization pass).
Upstream base : google/gemma-3-1b-it (Google LLC, Gemma Terms of Use)
NOESIS variant : NOESIS LoRA-merged Scenema student fine-tune.
PEFT merge_and_unload of NOESIS cinema r=8 +
autoif-v2 r=16 LoRA adapters into the BF16 base. Fully
merged single-shard model.safetensors; no residual
adapter layers.
NOESIS role : Recipe-validation companion / lightweight test
student for the primary
NOESIS-Gemma3-4B-Scenema-LTX2.3-BF16 student. Used
for Phase 1D scenema_adapter recipe smoke-tests
before scaling to 4B. Production Scenema text-encoder
path uses the 4B sibling; this 1B is the validation
+ experimentation slot.
Last updated : 2026-08-29
================================================================================ NOESIS Fine-Tune Lineage
google/gemma-3-1b-it
© Google LLC 2024-2025 (Gemma Terms of Use)
│
│ NOESIS cinema LoRA r=8 (60K pairs)
│ + autoif-v2 LoRA r=16
│ PEFT merge_and_unload → BF16
▼
NOESIS-Gemma3-1B-BF16 ← THIS BUNDLE
© AMAImedia 2026 (NOESIS fine-tune derivative)
Apache-2.0 for the NOESIS LoRA delta + packaging;
Gemma 3 base weights inherit Google Gemma Terms of Use.
The Scenema audio cross-attention pipeline needs 3840-d per-token text
embeddings (R-SCENEMA-KD-DIM-3840-NOT-5376). The 1B student's hidden
state is ~1152-d; the separately trained scenema_adapter.pt Linear
projection (1152 → 3840) bridges to the Scenema-DiT
audio_aggregate_embed layout — same pattern as the 4B sibling, just
with the smaller backbone for recipe validation.
================================================================================ Architecture
| Property | Value |
|---|---|
| Architecture | Gemma3ForConditionalGeneration (1B IT variant) |
| Total parameters | ~1B |
| hidden_size | ~1152 |
| num_hidden_layers | ~26 |
| num_attention_heads | 4 |
| num_key_value_heads | 1 (GQA 4:1) |
| vocab_size | 262,144 (Gemma 3 tokenizer) |
| Precision | BF16 (fully merged, no adapters) |
| File size | ~1.86 GB (single shard model.safetensors) |
| Scenema adapter target dim | 3840 (R-SCENEMA-KD-DIM-3840-NOT-5376) |
| Adapter projection | Linear(1152 → 3840), separately trained |
================================================================================ Bundle Contents
.
├── README.md
├── LICENSE # Apache-2.0 + NOESIS notice + Gemma Terms
├── config.json # Gemma3ForConditionalGeneration config
├── generation_config.json
├── chat_template.jinja
├── tokenizer.json
├── tokenizer_config.json
└── model.safetensors # BF16 weights (~1.86 GB, single shard)
================================================================================ NOESIS Scenema Student Family
| Bundle | Role | Size | Hidden |
|---|---|---|---|
| Gemma3-12B-it-Qat-Q4_0-Unquantized-BF16 | KD teacher | ~23.4 GB | 3840 (KD target dim) |
| NOESIS-Gemma3-4B-Scenema-LTX2.3-BF16 | Primary student | ~8.01 GB | 2560 → 3840 via adapter |
| NOESIS-Gemma3-1B-BF16 ← this bundle | Recipe-validation student | ~1.86 GB | ~1152 → 3840 via adapter |
| NOESIS-Gemma4-12B-it-Qat-Q4_0-Unquantized-BF16 | Route-2 G4→G3 adapter experiments | ~23.9 GB | 3840 (different distribution) |
================================================================================ NOESIS Sealed Rules
R-APACHE-CLEAN NOESIS LoRA delta + packaging released under Apache-2.0. Upstream Gemma 3 1B base weights remain under Google Gemma Terms of Use — redistribution must include attribution to Google and a copy of the Gemma Terms (see LICENSE NOTICE).
R-VENDORED-INTERNAL Internal vendor copy inside parent NOESIS-3.5B-A0.5B-DUBBING-FILM / LTX2.3-5B-Scenema-Audio-DiT-BF16.
R-SCENEMA-KD-DIM-3840-NOT-5376 (sealed 2026-05-21)
Scenema KD target dim = 3840 (Gemma 3 12B teacher hidden_size).
This 1B student's hidden states (~1152) are projected to 3840 via
a separately trained scenema_adapter.pt Linear head. Legacy 5376
is WRONG and must never be used.
R-SCENEMA-PER-TOKEN-NOT-MEAN-POOL (sealed 2026-05-21) Adapter output must be per-token [seq, 3840]. Mean-pool to [1, 3840] breaks Scenema-DiT audio cross-attention.
R-CINEMA-LORA-MERGED (sealed 2026-05-26) Cinema LoRA r=8 permanently merged into model.safetensors via PEFT merge_and_unload. No adapter files remain in this bundle.
R-AUTOIF-LORA-MERGED (sealed 2026-05-26) autoif-v2 LoRA r=16 permanently merged into model.safetensors via PEFT merge_and_unload. No adapter files remain in this bundle.
R-BF16-SOURCE-OF-TRUTH-NO-QUANT (sealed 2026-05-26) BF16 master, NOT dequantized from any downstream quant. All future re-quantization / distillation / fine-tuning must start here.
R-DUBBING-FILM-SCOPE (sealed 2026-04-29) NOESIS = professional audio dubbing. This bundle is the 1B-class recipe-validation companion for the Scenema audio text-encoder pipeline.
R-NOESIS-FINAL-ARTIFACT-PATHS (sealed 2026-05-27) Canonical path: models/llm/NOESIS-3.5B-A0.5B-DUBBING-FILM/ LTX2.3-5B-Scenema-Audio-DiT-BF16/NOESIS-Gemma3-1B-BF16/
R-NEVER-DELETE-WITHOUT-EXPLICIT-CONSENT (sealed 2026-05-21) MUST NOT be deleted without explicit operator instruction "удали " / "delete ". Authoritative BF16 master of the 1B Scenema recipe-validation student — loss requires re-running the LoRA merge from base.
================================================================================ Upstream Citation
Upstream base: https://huggingface.co/google/gemma-3-1b-it Gemma Terms: https://ai.google.dev/gemma/terms
================================================================================
NOESIS — Deterministic Hybrid Control Framework for Frozen Neural Operators (DHCF-FNO). Copyright (c) 2026 AMAImedia.com. All rights reserved. Gemma 3 1B base weights © 2024-2025 Google LLC, released under the Gemma Terms of Use (https://ai.google.dev/gemma/terms).
======================================== Scenema bridge v6 PAIRED (2026-06-06)
scenema_adapter_v3.pt = bridge_1b_v6.pt (ScenemaAdapterV3, 1152->188160, val_cos 0.8953). NOT weight-merged into model.safetensors (non-linear LayerNorm+GELU projector applied to OUTPUT hidden states, mathematically cannot fold into transformer weights). Loaded as paired runtime by scenema_wrapper.ScenemaEncoder.from_pretrained(this_dir): Gemma3-1B forward -> hidden[1152] -> ScenemaAdapterV3 -> Scenema-DiT [seq x 188160]. Plateau at ~0.895 (beats v4 baseline 0.88); en weakest 0.858, ko strong 0.965. Source: LORA/bridges_v6_2026-06-05/.