sys1-fast-encoder-v2

Part of sys1 — a small, open "System One" model: typed judgments (Choice/Score/Noul) with real calibrated probabilities, not generated text.

A sentence-transformers/all-MiniLM-L6-v2 encoder (22M params) with a one-neuron decision head, scored per candidate option in one batched forward pass. No chat template, no generation — this is what gets to millisecond-level decisions on CPU, unlike a causal chat model read via logits.

Usage

import sys
sys.path.insert(0, "src")  # or `pip install sys1`
from sys1 import Choice, FastEngine

engine = FastEngine("ashurath/sys1-fast-encoder-v2")  # or a local checkpoint dir
result = Choice(["billing", "technical support", "sales", "spam"]).ask(
    "Classify what this support ticket is about.",
    {"ticket": "My card was charged twice for the same order, please refund me."},
    engine,
)

Training

Trained on real labeled data pulled from public Hugging Face datasets (ticket routing/urgency, sentiment) plus the train-only splits of the same public benchmark used to compare TypeSafe's Jev and Convai's Laya (Luni/laya-jev-benchmark) — never touching that benchmark's held-out test data. See training/train_fast_encoder.py.

Measured results

	Own held-out eval	Public agent-trace benchmark	Public phishing benchmark
Accuracy	91.3%	64.4% (from 37.4% pre-domain-training)	56.5% (from 49.0%)
Brier score	0.119	0.263 (worse — see below)	0.370 (worse — see below)

Accuracy improved substantially on both public benchmarks after training on their train splits. Calibration (Brier score) got worse on both despite that — the model became more confident faster than it became more correct. This is an open issue, documented rather than hidden: see the project README for the full story.

Latency: ~35ms per decision end-to-end on a CPU laptop, no GPU.