YAML Metadata Warning:empty or missing yaml metadata in repo card
Check out the documentation for more information.
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-Qwopus3.5-4B-v3-Supervisor-LongCtx-BF16
Role: Long-context Supervisor — PRODUCTION (multi-segment QC, orchestration plan,
cross-stage review, batch best-of-N) for the dubbing pipeline.
From: base NOESIS-Qwopus3.5-4B-v3 + plain-LoRA nt346_longctx_qwopus4b (CCE, max_len 768,
completion-masked) merged. Trained on NF4 base, merged into BF16.
BF16 = PRIMARY. Siblings: -NF4 (6GB runtime), -Q8_0.gguf (laptop), LoRA adapter.
Test results (2026-06-17)
Supervisor eval — eval_longctx_supervisor_v1.py (12 tests, grammar-constrained)
| Quant | Size | Score |
|---|---|---|
| Q8_0 | 4.17 GB | 11/12 |
| IQ2_XXS | 1.43 GB | 8/12 (2-bit cliff: under-escalates on 248K vocab — not recommended) |
Only failure at 11/12 = L1_mixed (aggregates trunc+speed mix to reject vs retry;
fail-safe). Prior best long-ctx model (LFM2.5-7.5B) was 7/12.
Translation — FLORES devtest (chrF++/BLEU, n=20, no-think)
| Direction | chrF++ | BLEU |
|---|---|---|
| eng→rus | 51.5 | 25.6 |
| eng→cmn | 32.0 | 7.4 |
| AVG | 41.7 | 16.5 |
Comparison vs dedicated models (same FLORES n=20)
| Model | Supervisor-12 | Translate AVG chrF++/BLEU |
|---|---|---|
| This (4B-LongCtx) | 11/12 | 41.7 / 16.5 |
| base 4B (no LoRA) | 5/12 | 42.4 / 15.4 |
| Qwopus3.5-9B-Translate Q4 | 5/12 | 43.8 / 16.5 |
Takeaway: gains full supervision (+6 over base) with no loss of translation (parity with base; within ~2 chrF of the dedicated 9B translator at < half the params). Strong all-rounder.
Speed (RTX 3060 Laptop 6GB, GPU, 33/33 layers offloaded)
- Q8_0: gen 53.5 tok/s, prompt eval 366 tok/s. (vs 9B-Translate Q4: 49.1 / 307.)
Eval on GPU via standalone llama-completion.exe -ngl 99 (pip llama_cpp is CPU-only).
Written: 2026-06-17
- Downloads last month
- 18