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

NOESIS-Hy-MT2-7.5B-BF16

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
• Telegram: @AMAImediacom
• NOESIS version: v15.10
• Trained: 2026-06-05/06
• Last updated: 2026-06-06

⚠️ License regime — Tencent HY Community License Agreement. Three license gates apply: (1) Territory — EU excluded; (2) 100M MAU cap (separate Tencent license required above); (3) No-KD: outputs cannot be used to train other AI models. Full text in LICENSE.md.

---

Role in NOESIS pipeline

Primary translation backend for the NOESIS dubbing pipeline (Stage 3 source→target compress translate). 33 mainland + 5 ethnic + Cantonese languages.

This bundle = BF16 merged checkpoint of tencent/Hy-MT2-7.5B + NOESIS SFT-LoRA adapter (nt312_sft_hymt2_7b) trained on the NOESIS dubbing corpus. Format: bf16 dense (master). Disk ≈ 14.0 GB, VRAM peak load ≈ 14.0 GB on RTX 3060.

---

NOESIS A/B test results — 2026-06-06

Evaluation harness: scripts/nt318_eval_ab.py (chat-format prompt + <|im_end|> stop + held-out 30-sample slice of dub_subscene_MERGED.jsonl, filtered to common target langs).

Metric	Upstream NF4 (baseline)	NOESIS SFT (this)	Δ
garbage% (n=30)	16.7%	0.0%	✅ -16.7 pp
iso_fit (dubbing budget)	0.502	0.549	✅ +0.047
overlap vs gold (Jaccard)	0.202	0.226	✅ +0.024

Verdict: ✅ no regression + measurable quality improvement on every metric.

Real example outputs

Source	Upstream NF4	NOESIS SFT
Свртете се на десно.	`Translation: "Turn right."	`
Можам ли да ги видам?	Конеч语可以吗?Can we see them?	I can see them.

Old/new sample dumps: logs_heal/ab2/{old,new}_hymt2_7b.json.

---

Bundle inventory

File	Description
model.safetensors	weights (BF16, single shard)
config.json	model architecture configuration
generation_config.json	default decoding params
tokenizer.json / tokenizer_config.json	tokenizer
chat_template.jinja	chat template (im_start / im_end markers)
NOESIS_MERGE_MANIFEST.json	NOESIS provenance (base, adapter, created-at)
README.md	this file
LICENSE.md	NOESIS provenance + upstream license terms

---

Training details

Field	Value
Method	SFT (Supervised Fine-Tuning) with QLoRA
Trainer	scripts/nt312_train_sft_lora.py
LoRA rank / alpha	16 / 32
LoRA targets	down_proj, q_proj, v_proj (lean)
Max steps	500, save every 50, --resume capable
LR scheduler	warmup_stable_decay (WSD, MiniCPM-style)
Optimizer	AdamW 8-bit (paged)
Adapter	LORA/nt312_sft_hymt2_7b/adapter
Dataset	LORA/Hy-MT2-SFT-100k.jsonl
Sealed rule	R-SEALED-LORA-RECIPE-V3-NF4

---

Quick Start

Load

from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

path = "NOESIS-Hy-MT2-7.5B-BF16"
tok = AutoTokenizer.from_pretrained(path, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(
    path, torch_dtype=torch.bfloat16, device_map={"": 0},
    trust_remote_code=True,
)

Translate (chat format — MANDATORY for correct stop)

src_lang, tgt_lang, src = "en", "ru", "Get out of the city, now!"
prompt = (
    f"<|im_start|>user\n"
    f"Переведи реплику дубляжа {src_lang}->{tgt_lang}, только перевод.\n"
    f"SOURCE ({src_lang}): {src}<|im_end|>\n"
    f"<|im_start|>assistant\n"
)
stop_ids = [tok.eos_token_id, tok.convert_tokens_to_ids("<|im_end|>")]
ids = tok(prompt, return_tensors="pt").input_ids.cuda()
g = model.generate(
    ids, max_new_tokens=64, do_sample=False,
    repetition_penalty=1.2,
    eos_token_id=stop_ids, pad_token_id=tok.eos_token_id,
)
print(tok.decode(g[0][ids.shape[1]:], skip_special_tokens=True))
# → "Уберись из города! Немедленно."

Convert via convert_hf_to_gguf.py (b8808 patched) → llama-quantize q5_k_m/q8_0 (b9523).

---

NOESIS sealed rules

Rule	Summary
R-MT-PRIMARY-7B-NF4-FALLBACK-1.8B	Hy-MT2 7.5B = primary translator; 1.8B = low-VRAM fallback
R-MT-OWN-NF4-NOT-GGUF	Translator NF4 is OWN bnb quant, not GGUF (GGUF is a separate dense artifact)
R-SEALED-LORA-RECIPE-V3-NF4	Canonical SFT-LoRA recipe (rank 16, lean targets, WSD)
R-HF-NAMING-QUANT-ONLY-PRESERVE-UPSTREAM	Naming convention for trained derivatives
R-NEVER-DELETE-WITHOUT-EXPLICIT-CONSENT	Bundle must not be deleted without explicit operator instruction

---

Upstream

• Base model: tencent/Hy-MT2-7.5B
• License: Tencent HY Community License Agreement — see LICENSE.md for the full text and NOESIS compliance notes.
• Training corpus: internal NOESIS dubbing dataset (translation pairs with phoneme budgets + isochrony targets).

NOESIS provenance metadata, bundle inventory, sealed-rule annotations, and DHCF-FNO integration notes © AMAImedia 2026 (NOESIS DHCF-FNO project).

MT benchmark — FLORES-200 devtest (2026-06-17)

Real eval (not smoke): n=100 × 4 directions (eng↔rus, eng↔cmn), GPU via resident llama-server -ngl 99. Primary metric COMET (wmt22-comet-da, neural — how "best translator" is judged), plus chrF++ / BLEU / length-ratio. Each model prompted in its own native format (MT2 = dubbing ChatML "SOURCE (lang): … Только перевод"; 9B = ChatML + no-think). Data + COMET checkpoint: D:/models/by_expert/07_MT_TRANSLATION.

Model	Size	COMET avg	chrF++	BLEU	gen tok/s
Qwopus3.5-9B-Translate Q4	5.24 GB	0.8870	50.7	22.5	49
NOESIS-Hy-MT2-7.5B Q5	5.0 GB	0.8709	46.2	21.4	52
NOESIS-Hy-MT2-1.8B Q8	1.78 GB	0.8481	43.9	19.1	121

Per-direction COMET — 9B-Translate wins all 4 (eng-rus .902 / eng-cmn .897 / rus-eng .872 / cmn-eng .877); MT2-7.5B 2nd, MT2-1.8B 3rd.

Notes:

• MT2 is a dubbing translator (isochrony): its outputs are shorter (len_ratio ~0.87-0.89 vs 9B ~1.0) because it compresses to fit speech slots → lower chrF on literal FLORES news. FLORES does NOT measure MT2's slot-fit strength, so it under-rates MT2 for its actual job.
• 1.8B→7.5B degradation: COMET +0.023, chrF +2.3, BLEU +2.3 — modest; 1.8B is 2.4× faster and 2.8× smaller (good lightweight tradeoff).
• BLEU for eng-cmn is low for all (Chinese needs char-tokenization); use chrF++/COMET there.