CoolFace
Modelpublic

AMAImedia/NOESIS-Llama-1B-MiniCPM5-BF16

sourceHugging Faceapache-2.0updated 28d agoView on Hugging Face
0likes879downloads
Model Card

NOESIS / AMAImedia

Released as part of the NOESIS Professional Multilingual Dubbing Automation Platform (framework: DHCF-FNO — Deterministic Hybrid Control Framework for Frozen Neural Operators).

================================================================================ NOESIS-Llama-1B-MiniCPM5-Orchestrator-Director-Supervisor-BF16 -- NOESIS Bundle README ================================================================================

NT-325 SFT director merge — peft.merge_and_unload of the NOESIS director-SFT LoRA (LORA/nt325_sup_llama1b_minicpm5/adapter) over the NOESIS Llama-1B-MiniCPM5 Supervisor BF16 base. Role: director / supervisor / orchestrator of the NOESIS dubbing pipeline — dense Llama-arch small director, 128 K context, terse decision-only output, no <think>.

NOESIS provenance ----------------- Bundle : NOESIS-Llama-1B-MiniCPM5-Orchestrator-Director-Supervisor-BF16 Parent bundle : NOESIS-3.5B-A0.5B-DUBBING-FILM Upstream chain : openbmb/MiniCPM5-1B (Apache-2.0) → NOESIS-Llama-1B-MiniCPM5-Supervisor-BF16 → NT-325 director-SFT LoRA merge → this bundle License : Apache License 2.0 (end-to-end) NOESIS variant : BF16 merged director (single dense Llama shard, ~2.16 GB). NT-325 director-SFT LoRA merged into the Llama-1B-MiniCPM5 Supervisor base via peft.merge_and_unload. NOESIS role : 1 B-tier director / supervisor / orchestrator with 128 K context (full cinema reel fits in one context). Sibling to the 0.8 B Qwopus director; both consume the same NOESIS-SUPERVISOR-DIRECTOR-FINAL-v3-97k.jsonl curriculum. Last updated : 2026-08-29

================================================================================ NOESIS director ladder (small → large) ================================================================================

TierBundleRole
0.8 B dense../NOESIS-Qwopus3.5-0.8B-v3-Orchestrator-Director-Supervisor-BF16Fast small director (sibling)
1.0 B denseNOESIS-Llama-1B-MiniCPM5-Orchestrator-Director-Supervisor-BF16 ← this bundleDense Llama-arch small director, 128 K context
7.5 B MoE A1B../NOESIS-LFM2.5-7.5B-A1B-Orchestrator-Director-Supervisor-BF16Hybrid conv+attn MoE director
7.5 B MoE A2.5B../NOESIS-Mellum2-7.5B-A2.5B-Orchestrator-Director-Supervisor-BF16QWEN3MOE director

================================================================================ Architecture / Files ================================================================================

PropertyValue
ArchitectureLlamaForCausalLM (standard, R-LLAMA-ARCH-STANDARD)
Total params~1.08 B
Layers24
Attention headsGQA 16 Q / 2 KV
Context length131 072 (128 K) — full cinema reel fits in one context
PrecisionBF16 (single shard)
Weights filemodel.safetensors (~2.16 GB)
TokenizerMiniCPM5 tokenizer (tokenizer.json + tokenizer_config.json)
Chat templatechat_template.jinja (director / no-think)
Deploy quantGGUF Q8_0 (sister deploy file, A/B clean)
.
├── README.md
├── LICENSE                              # Apache 2.0 + NOESIS notice
├── model.safetensors                    # BF16 dense Llama (~2.16 GB)
├── config.json
├── generation_config.json
├── tokenizer.json
├── tokenizer_config.json
├── chat_template.jinja
└── NOESIS_MERGE_MANIFEST.json           # NT-325 merge provenance

================================================================================ 3-way A/B — OLD upstream vs NEW NT-325 SFT vs GGUF Q8_0 (5 supervisor prompts) ================================================================================

Operator-verified 2026-06-06:

VariantGarbageSpeedQuality
OLD upstream (no NT-325 SFT)5 / 5 (all <think>)15.1 tok/s❌ everything in reasoning trace
NEW NT-325 SFT (this bundle)0 / 5 ✅~17 tok/s⚠️ no <think>, but some answers weak (compress/expand direction sometimes wrong)
GGUF Q8_0 (sister deploy file)0 / 5 ✅—⚠️ prompt-echo, but Paris, works

→ NT-325 SFT removed the <think> leak ✅. Direction accuracy on isochrony decisions is below the 0.8 B Qwopus sibling — for hard routing / isochrony cases prefer the Qwopus 0.8 B director (sharper decisions); use this 1 B Llama bundle when long context (>32 K, up to 128 K) is required (full reel scope, multi-speaker QC).

================================================================================ How it was built — NT-325 director SFT merge ================================================================================

  1. 1.Base — NOESIS-Llama-1B-MiniCPM5-Supervisor-BF16 (Supervisor BF16 source, fine-tuned from openbmb/MiniCPM5-1B).
  1. 1.Director-SFT LoRA — LORA/nt325_sup_llama1b_minicpm5/adapter. Trained on LORA/NOESIS-SUPERVISOR-DIRECTOR-FINAL-v3-97k.jsonl (same curriculum as the 0.8 B / 7.5 B directors): routing / isochrony / gate decisions, canon-RZ22 stage transitions, no-<think> enforcement.
  1. 1.Merge — peft.merge_and_unload → BF16 single shard (scripts/nt325_merge_director_bf16.py, 2026-06-06 21:29:21).
  1. 1.GGUF Q8_0 deploy — sister deploy file. A/B 0 / 5 garbage, prompt-echo visible but Paris probe passes.

================================================================================ VRAM and runtime ================================================================================

SetupVRAM
BF16 native, no offload~2.6 GB
BF16 + KV @ 4 K ctx~3.2 GB
BF16 + KV @ 32 K ctx (typical cinema reel)~4.5 GB
RTX 3060 6 GB✅ comfortable with sequential-swap policy

================================================================================ NOESIS Sealed Rules ================================================================================

R-APACHE-CLEAN End-to-end Apache 2.0 lineage (MiniCPM5-1B → NOESIS Supervisor BF16 → NT-325 director SFT merge). Full upstream license text + NOTICE in LICENSE.

R-SUPERVISOR-ALWAYS-CONNECTED Director / supervisor wired into the canon-RZ22 dub pipeline via demo_server/qwopus_supervisor.py (NEVER delete / orphan).

R-SFT-DIRECTOR-NO-THINK (pattern) Production verdicts are flat, terse, decision-only. No <think> / CoT leakage. enable_thinking=False in production.

R-LLAMA-ARCH-STANDARD Standard LlamaForCausalLM — no custom kernels, no model-code fork. HF-loadable out of the box.

R-TRAIN-CKPT-50-RESUMABLE NT-325 director SFT training: checkpoints every 50 steps, spot-instance resumable.

R-VENDORED-INTERNAL Internal vendor copy inside parent NOESIS-3.5B-A0.5B-DUBBING-FILM.

R-DUBBING-FILM-SCOPE (sealed 2026-04-29) NOESIS = professional audio dubbing. 1 B-tier director / supervisor of the canon-RZ22 dub pipeline. 128 K context for long-form multi-speaker reels.

R-NOESIS-FINAL-ARTIFACT-PATHS (sealed 2026-05-27) Canonical path: models/llm/NOESIS-3.5B-A0.5B-DUBBING-FILM/ NOESIS-Llama-1B-MiniCPM5-Orchestrator-Director-Supervisor-BF16/

R-NEVER-DELETE-WITHOUT-EXPLICIT-CONSENT (sealed 2026-05-21) MUST NOT be deleted without explicit operator instruction "удали <path>" / "delete <path>". 128 K-context director — long-reel QC role can't be filled by the 0.8 B Qwopus sibling.

================================================================================ Upstream Citation ================================================================================

Upstream model : https://huggingface.co/openbmb/MiniCPM5-1B License : Apache License 2.0 NT-325 SFT adapter source : LORA/nt325supllama1b_minicpm5/adapter Training corpus : LORA/NOESIS-SUPERVISOR-DIRECTOR-FINAL-v3-97k.jsonl

bibtex
@article{minicpm4,
  title  = {MiniCPM4: Ultra-efficient LLMs on end devices},
  author = {MiniCPM Team},
  journal= {arXiv preprint arXiv:2506.07900},
  year   = {2025}
}

================================================================================

NOESIS — Deterministic Hybrid Control Framework for Frozen Neural Operators (DHCF-FNO). Copyright (c) 2026 AMAImedia.com. All rights reserved. MiniCPM5-1B base weights © 2025 OpenBMB / MiniCPM Team, released under the Apache License 2.0. NOESIS Supervisor BF16 fine-tune + NT-325 director SFT merge + provenance © AMAImedia 2026 (NOESIS DHCF-FNO project, Apache-2.0).