nev-lite-systemone
nev-lite: 11.8M-parameter static-embedding decision head that scores fixed option sets with no transformer pass at inference.
Overview
nev-lite is the static, no-attention decision head of the nev / System One engine, published on Hugging Face. It answers typed questions over a known, fixed option set without a transformer forward pass: state tokens are looked up in a learned embedding table, pooled and projected, option vectors are pooled once and cached, and a small interaction MLP scores the state against each option. Per-request cost depends only on state length, not the number of options.
The model has about 11.8M parameters, dominated by a 30522x384 embedding table seeded from all-MiniLM-L6-v2. It supports choice (benchmarked on Banking77, CLINC150, AG News, emotion, TREC), score (SST-5) and noul (BoolQ, SST-2, prompt-injection detection). Outputs are margins rather than probabilities; a temperature or Platt scaler must be fitted separately. It is aimed at high-volume, low-cost classification where options rarely change.
What the decision model decides
| Type | Question | Answers |
|---|---|---|
| Choice | Pick one of K fixed labels (e.g. intent) | — |
| Score | Pick an ordered level (e.g. sentiment very negative to very positive) | — |
| Yes / no | Is this true (e.g. is this a prompt injection)? | true · false |
How it works
- Cache pooled vectors for the fixed option set.
- Tokenise and pool the state with the static embedding table.
- Score each option with the interaction MLP.
- Apply a fitted temperature or Platt scaler for probabilities.
Strengths
- Very low cost per decision; no attention at inference.
- Cost independent of the number of options.
- Apache-2.0 weights.
Caveats
- Only for known, fixed option sets.
- Margins are uncalibrated until a scaler is fitted.
- Bag-of-embeddings pooling loses word order.
Getting started
Download nev_lite.pt, option_cache.pt and config from https://huggingface.co/shreyanbr/nev-lite-systemone; source is at github.com/Shreyan1/nev-system-one.At a glance
- Type
- Model
- Status
- research
- Language
- Python
- Runtime
- PyTorch
- Has code
- yes
- License
- Apache-2.0
Use cases and integrations
How curated lists describe it
No upstream description was published.
wh000wh000/awesome-jev-live
Frequently asked questions
Why no attention?
It trades contextual understanding for cost: a decision is a lookup, a pool and a few dot products.
Are the outputs probabilities?
No, they are margins; fit a temperature or Platt scaler first.
Can it handle new options at request time?
It is built for fixed option sets whose vectors are cached.
Related projects
Japanese System One decision model on ModernBERT-ja (314.6M) returning typed choice, score and bool distributions with no generation.
ProjectbetaPythonOEV: 184M-parameter decision engine that scores options in one pass, reporting 0.7705 on typed-decisions and 22 ms per question.
ProjectresearchPythonFastino's 340M Apache-2.0 decision classifier that scores caller-defined labels in one forward pass for routing, intent and scoring.
ModelproductionPythonOpen multilingual System 1 decision engine built on Laya checkpoints with a router that picks the best checkpoint per request.
ModelbetaPython- nirnay★ 9
NIRNAY 450M: open Apache-2.0 decision model fine-tuned from Laya that reports 87.9% on Banking77 with fitted ECE 0.045.
ModelresearchPython Superagent's Apache-2.0 security decision model from AutoJev-27B, scoring options for injection, tool calls, code and alerts.
Modelbeta