JevSpan

It splits text into sentences and clauses, offers every candidate window as an option in a Jev Choice question such as which window is exactly one complete drug, verifies each nominated window with labels including type, none, mixed and partial, then lets Jev pick the exact boundary from trimmed and extended variants. Every entity returns a probability and a trace.

73.7 strict F1 averaged across 12 public Chinese and English benchmarks, including the five CrossNER domains, with no labelled training data and Jev 1.13. On the seven datasets no compared method was trained on, it leads Baidu UIE and GLiNER by roughly 21 to 28 points, and on the five CrossNER domains it is about 15 points ahead of NuNER-Zero.

A schema is a JSON object where each type needs a description, with title, examples and counter_examples optional, and a shorthand form like {"drug": "a medicine's name"} also works. Two further options are include_brackets, which keeps the surrounding 《》 for book and film titles, and min_chars, default 2, which you set to 1 for single-character entities.

The boundary convention. On a small e-commerce set, adding one sentence saying a model name excludes the product category and version suffix and that a price excludes the word about raised strict F1 from 0.778 to 0.945. Ready-made schemas ship in eval/schemas, src/jevspan/presets and one per benchmark under bench/schemas.

Python 3.12 or newer, pip install jevspan and TYPESAFE_API_KEY from the environment or a .env in the current directory, then jevspan on a string or jevspan-web on http://127.0.0.1:47321; from source use uv sync. Each entity carries text, label, start and end character offsets, Jev's score and its source, and --trace shows every segment, candidate and verdict while the web UI draws the full decision tree. Every Jev answer is cached in .cache/jev.sqlite, so repeating a run is free.

Every method was zero-shot and saw the same 200 test sentences per dataset at a fixed seed and the same type descriptions, and all compared models are under 30B parameters. Measured September 2026 with jev-1.13.0, JevSpan leads Qwen3.8-27B by 10 points on MSRA, 6.5 on Weibo, 6.4 on CrossNER politics and 3.9 on CLUENER, while trailing by 4.2 on CoNLL and 8.4 on MIT Restaurant.

The splitter understands CJK punctuation, full-width brackets, book-title marks and space-delimited Latin text, so the same library handles a Chinese social post and an English news paragraph. include_brackets is the schema switch for keeping the surrounding 《》 on book and film titles, and min_chars defaults to 2 but can be set to 1 for single-character entities.