classone-gemma4-e2b — ClassOne System 1 Decision Model

devops-thiago/classone-gemma4-e2b is an open-source System 1 decision model built on google/gemma-4-E2B-it using the ClassOne (Classifier One) architecture. The merged standalone backbone ships directly in this repository — it loads as a single model, with no adapter and no separate base-model download.

Instead of generating text token by token, ClassOne evaluates structured decisions in a single forward pass, returning typed, calibrated outputs with zero decoding overhead.

Self-Measured Benchmark Results

Self-Measured Evaluation Provenance: Executed locally on 2x NVIDIA RTX 5060 Ti 16 GB (Date: 2026-11-13, Commit: c375207). See Official Results (Benchmark Heaven) below for third-party sealed evaluation.

Tokenizer Compatibility Note (commit 86988ca / SDK v0.2.0–v0.2.3): The standalone weights in this repository were trained with custom delimiter tokens (<|state_start|>=262144..262155) serialized in tokenizer.json. Later SDK releases (v0.2.0–v0.2.3) prioritize Gemma 4 <unused0>.. reserved tokens when present; with native <|state_start|>=262144 formatting, these weights + rolled-back classone_heads.pt (F6A6E423...) achieve 69.3% (160/231) on JevBench (whereas default <unused0> formatting scores 58.4% / 135/231).

1. JevBench Public Multi-Tier Suite (231 Public Tasks, Self-Measured Standalone)

Self-measured on 2x NVIDIA RTX 5060 Ti 16 GB · Date: 2026-11-13 · Commit: c375207

Evaluation Regime	Overall (231)	Easy (48)	Original (72)	Hard (111)
Current Hub Weights + Rolled-Back Heads (`F6A6E423...`, Native `<\|state_start\|>`)	69.3% (160/231)	93.8% (45/48)	88.9% (64/72)	45.9% (51/111)
Original Milestone Snapshot (`best_composite_70pct_milestone`, 2026-10-06)	70.1% (162/231)	100.0% (48/48, 0.0064 ECE)	90.3% (65/72, 0.0972 ECE)	44.1% (49/111)
Current Hub Weights under SDK `v0.2.0–v0.2.3` (`<unused0>`)	58.4% (135/231)	91.7% (44/48)	70.8% (51/72)	36.0% (40/111)

2. RLCDAlignBench Alignment & Safety Suite (100 Instances, 2026-10-06 Snapshot)

Self-measured on 2x NVIDIA RTX 5060 Ti 16 GB · Date: 2026-11-13 · Commit: c375207

Failure Mode / Axis	Samples (N)	AUROC	Accuracy (%)	ECE	Latency (p50)
Power Seeking	6	0.889	83.3%	0.2575	209.6 ms
Honesty (Deception)	11	0.500	72.7%	0.3906	216.8 ms
Concealing Uncertainty	14	0.673	71.4%	0.1800	127.8 ms
Refusal (Jailbreaks)	11	0.500	63.6%	0.0974	267.5 ms
Faithfulness	9	0.700	55.6%	0.2513	200.0 ms
Bias	9	0.525	55.6%	0.2141	217.9 ms
Overall Balanced Accuracy	100	0.505	56.2%	0.1944	199.2 ms

3. Tool-Routing & Safety Guardrail Suites (Self-Measured)

Self-measured on 2x NVIDIA RTX 5060 Ti 16 GB · Date: 2026-11-13 · Commit: c375207

Benchmark Suite	Split / Samples	Accuracy	Secondary Metric	Calibration ECE	Latency (p50)
WildGuardTest (`walledai/WildGuardTest`)	`test` (100 stratified)	72.00% (72.04% BalAcc)	0.7993 AUROC (78.6% Vanilla, 72.9% Benign Spec)	0.1597	43.5 ms
Deepset Prompt Injections (`deepset/prompt-injections`)	`test` (116 full)	50.86% (59/116)	0.6214 AUROC (75.0% Injection Recall)	0.3412	41.6 ms
NVIDIA When2Call (`nvidia/When2Call`)	`mcq` (100 stratified)	35.00%	85.3% `tool_call` (29/34)	0.5376	106.8 ms

Official Results (Benchmark Heaven)

Official third-party JevBench leaderboards are hosted at Benchmark Heaven (checked 2026-11-13, JevBench v1.8.10 / method v1.7, published 2026-11-12; see the live entry for classone-qwen3.5-9b). Official Benchmark Heaven scores differ from local self-measurements because the official harness evaluates private sealed items in addition to the public suite, applies a public/sealed generalization gap penalty, enforces a strict 10-second cutoff rule, and measures latency/cost on server-grade GPUs rather than local workstation hardware (2x NVIDIA RTX 5060 Ti 16 GB).

Edge vs Cloud Latency (Self-Measured Comparison)

Self-measured on 2x NVIDIA RTX 5060 Ti 16 GB · Date: 2026-11-13 · Commit: c375207

Measured against a remote cloud LLM decision API round-trip:

ClassOne (Local RTX 5060 Ti): 52.49 ms mean latency (19.1 req/s, $0.00 inference cost, 100% private)
Cloud API Baseline: 329.90 ms mean latency (3.0 req/s)
Edge Speedup: 6.3× faster than cloud API round-trip latency

Decision Primitives

Noul — Boolean check returning a calibrated probability P(true) ∈ [0, 1]
Choice — Categorical selection over 2–255 dynamic options with full probability distribution
Score — Continuous ordinal rubric rating over 2–10 levels (expected value)

All outputs are calibrated with a combined focal NLL (γ=2.0) + normalized Brier + contrastive margin loss and post-hoc temperature scaling.

Quickstart

pip install classone

import torch
from huggingface_hub import hf_hub_download
from transformers import AutoTokenizer

from classone.modeling.modeling_classone import ClassOneModel
from classone.schemas import NoulQuestion, ChoiceQuestion, ScoreQuestion
from classone.tokenizer import ClassOnePromptBuilder

REPO_ID = "devops-thiago/classone-gemma4-e2b"

# 1. Load the ClassOne model (weights + tokenizer are fully self-contained here)
tokenizer = AutoTokenizer.from_pretrained(REPO_ID)
builder = ClassOnePromptBuilder(tokenizer)
model = ClassOneModel.from_backbone(
    base_model_name_or_path=REPO_ID,
    tokenizer=tokenizer,
    device="cuda",
    torch_dtype=torch.float16,
)

# 2. Load the trained decision heads
heads = torch.load(hf_hub_download(REPO_ID, "classone_heads.pt"), map_location="cuda")
model.noul_head.load_state_dict(heads["noul_head"], strict=False)
model.choice_head.load_state_dict(heads["choice_head"], strict=False)
model.score_head.load_state_dict(heads["score_head"], strict=False)
model.eval()

# 3. Pack state + questions and run a single forward pass
packed = builder.pack(
    state={"customer": "Alex", "message": "I was charged twice for order #123."},
    questions={
        "refund": NoulQuestion(instructions="Is the user requesting a refund?"),
        "dept":   ChoiceQuestion(
                      instructions="Route to team:",
                      criteria={"billing": "Payment issues", "tech": "Technical bugs"}
                  ),
        "anger":  ScoreQuestion(
                      instructions="Dissatisfaction level:",
                      criteria=["satisfied", "neutral", "dissatisfied", "churning"]
                  ),
    }
)
results = model.evaluate_packed(packed)

print("Refund P(true):", results["refund"].noul)
print("Department:    ", results["dept"].choice, "—", results["dept"].probabilities)
print("Anger score:   ", results["anger"].score)

Repository Files

File	Description
`model.safetensors` (or sharded)	Merged ClassOne standalone backbone weights
`config.json`	Model configuration
`tokenizer.json`, `tokenizer_config.json`	Tokenizer with ClassOne delimiter tokens
`classone_heads.pt`	Trained Noul / Choice / Score head weights + calibrated temperatures

Certified classone_heads.pt SHA256: F6A6E42365158D99807B8995C800429A0BD2B12755EDC1DEC5823400D35F337D

Citation

@misc{classone2026,
  title={ClassOne: A Fast Single-Pass Decision Architecture for Language Models},
  author={Thiago Gonzaga},
  year={2026},
  url={https://github.com/devops-thiago/class-one},
}