JevmaxxxinG — 7-Member Specialist Ensemble (~2.65B total)

Not a single model. Seven fine-tuned encoder decision members — 4× ModernBERT-large (421M each) + 3× mmBERT-base (322M each), ≈2.65B parameters total — each an encoder + 2-layer head (no causal LM, no token generation), plus a domain router that forwards each request to the relevant specialist. System-1 decisions with calibrated probabilities, p50 15–25 ms, p95 ≤ 100 ms, 5.4 GB bf16 resident.

Load & run

Requires laya (runtime) — each member is a valid laya checkpoint:

pip install laya
from jevmaxxxinG import Ensemble

ens = Ensemble.from_pretrained("chortochki/JevmaxxxinG", device="cuda")   # ~5.4 GB bf16, resident
out = ens.decide("Хочу вернуть товар", {"intent": {
    "type": "choice", "instructions": "intent?",
    "criteria": {"return_refund": None, "track_order": None, "general": None}}})

out["answers"]["intent"]["choice"]         # 'return_refund'
out["answers"]["intent"]["probabilities"]  # {'return_refund': 0.93, ...}
out["answers"]["intent"]["confidence"]     # 0.91
out["routing"]                             # {'member': 'ru_rlcd', 'domain': 'auto'}

decide() / predict() take the exact laya.Agent.predict shape ({qid: {type, instructions, criteria}}) and return {model, answers, usage, routing}. The router picks the specialist from the request's domain (explicit domain=, Russian script → ru_rlcd, typed-decisions id-set → td, else the domain default). ens.load_all() loads all seven at once (≈394 s once, then resident).

Benchmarks — common protocols, same protocol for both

The defensible comparison: seven domains where both this ensemble and the official laya.Router were measured on the identical cases. OpenJev 4B is only comparable where its published protocol overlaps (JevBench + NLI) — it is not averaged into the mean.

Domain (held-out)	n	JevmaxxxinG	laya.Router (measured)	OpenJev 4B (pub)
JevBench (all)	1,107	0.918	0.589	0.814
typed-decisions (2,000 decisions)	2,000	0.769	0.769	—
NLI (MNLI-matched)	2,500	0.705	0.559	—
MASSIVE en-US intent	2,974	0.657	0.658	—
MASSIVE ru-RU intent	2,974	0.882	0.405	—
FOLIO formal logic	203	0.468	0.434	—
model-collapse reasoning	1,682	0.677	0.443	—
Mean over these 7		0.725	0.551	(not averaged — 2/7 comparable)

JevBench tiers (p14c member): easy 1.000 / standard 0.986 / hard 0.838 (OpenJev pub: 1.000 / 0.986 / 0.622). NLI detail (p12, openjev protocol): MNLI m/mm 0.843/0.853 (OpenJev 0.896/0.899), ANLI r1/r2/r3 0.521/0.407/0.416 (0.780/0.665/0.627), WANLI 0.653, SciTail 0.837, ConTRoL 0.517 — we trail OpenJev on domain-shifted NLI.

Zero-shot general classification (base member, all-domain-labels protocol)

Set	options	JevmaxxxinG	Jev 1.13 (pub)
AG News	4	0.938	0.950
DAIR Emotion	6	0.585	0.595
Banking77	77	0.393	0.425

Banking77 (77-way) is the many-option stress test — both models are weak there. The base member is convaiinnovations/laya base.

Calibration

MNLI-matched: ECE 0.023 (m) / 0.014 (mm), Brier 0.233, NLL 0.426, p50 21 ms. Order-invariance: 200/200 label-stable, max probability change 0.0 (bit-exact).

Speed & memory

• 5.4 GB bf16 resident (fp32: 10.70 GB) — fits an 8 GB GPU.
• p50 15–25 ms, p95 ≤ 100 ms per request (7-member, GPU-resident).
• Not an LLM: no causal head, no KV cache, no continuous batching → vLLM / llama.cpp do not apply. Runtimes: laya (in-process), ONNX Runtime, Triton.

Runtimes

• torch / laya (default): Ensemble.from_pretrained(...) — each member is a valid laya checkpoint (model.safetensors + rl_agent_config.json + tokenizer/).
• ONNX Runtime (no torch): each member exported to dynamic-shape ONNX under onnx/ (round-trip max logit diff 4e-6, prob diff 1e-7).
• Triton: jevmaxxxinG/serve_triton.py generates a Triton model repo from onnx/.

Limitations (measured / structural)

• This is an ensemble of per-domain specialists, not a generalist. Each member is fine-tuned on one domain and the router selects by domain signal; the mean above reflects that best-of routing, not a single model's general ability.
• The router is domain-routed, not a learned general router. It keys on the request domain (typed-decisions id-set, Russian script, explicit domain=); it was tuned against the benchmark routing signal and has no hold-out routing evaluation on unseen domains.
• NLI under domain shift trails OpenJev (ANLI/WANLI/ConTRoL 0.41–0.52 vs 0.63–0.77).
• FOLIO 0.468 is weak for formal logic.
• Languages: English-centric + one Russian member; multilingual is a separate checkpoint, not in this ensemble.
• Response-level factuality (LLM-AggreFact, RAGTruth, HaluBench, IFEval) is judged under a different, LLM-judged protocol than option selection; this ensemble is not evaluated on those sets, so we do not claim a number there.

What's inside

members/{p14c,p12,p8a,td,p16,ru_rlcd,base}/   model.safetensors + rl_agent_config.json + tokenizer/ + encoder/config.json
onnx/{...}/                                    dynamic-shape ONNX (fp32) per member
router/router.json                             learned domain router (bench_table + logistic)
jevmaxxxinG/                                   python package (Ensemble.from_pretrained / decide)
benchmarks.png                                 infographic

Built with laya (convaiinnovations) on top of answerdotai/ModernBERT-large / jhu-clsp/mmBERT-base. No teacher distillation: cross-entropy + judge-RLCD (opencode-cloud).