Open-type entities

Entity types named at run time: Nlp.Entities(text, "person", "organization"), with a threshold Nlp.Entities(text, ["contract party"], 0.1f), or OpenEntityRecognizer.Recognize(doc, types) on a document you tokenized.

Model

• The encoder reads the text once. Candidate spans (up to 12 tokens, flat in the output) get a vector from the span's first, last and mean token vectors and its width; every type name is run through the same encoder and mean-pooled and passed through a feed-forward layer to one vector. A span scores a type with a dot product of the two (scaled by 1/sqrt(width), plus a learned offset) in a shared 128-wide space, and a sigmoid turns it into a probability (Zaratiana et al. 2023, sections 3.2 to 3.4; the paper only).
• Training uses binary cross-entropy per token-type cell. A type an example lists with no spans is a negative; a type it does not list teaches nothing, since one data set may label "person" and not "actor". Up to negativeTypes negatives are drawn per example.
• Prediction keeps spans at or above the threshold (0.5 by default, per call with threshold), and where two overlap the more probable wins. The label is the type string as you gave it. Type vectors are cached, so a fixed list of types costs one encoder pass per type, once.
• Types: 1 to 256 distinct strings of at most 200 characters, in any language the encoder reads (the shipped head is trained on English).

Happy path

Nlp.Entities(string text, params string[] types) finds the registered open-type pipeline (ModelPackages.RegisterOpenEntities, called by the model package) and throws, naming RecursiveAI.OneNlp.Models.Multilingual, when there is none. Nothing is downloaded. A pipeline's Recognize is thread-safe.

The default threshold of 0.5 was the best on the Few-NERD and UNER dev sets, and it suits common type names: "organization" finds "Acme Corp." (0.76) and "Beta LLC" (0.80). An unusual name such as "contract party" scores far lower on the same text (0.15 to 0.17), so it returns nothing at 0.5. Pass a lower threshold to Nlp.Entities(text, types, threshold), and check the results, or use a type name the model has seen. To hold the recognizer yourself, MultilingualModel.LoadOpenEntities() returns a pipeline whose Components include an IOpenEntityRecognizer; its Recognize(doc, types, threshold) takes the same threshold (null for the component's default).

Data and limits

The recipe (recipes/openner) trains on Few-NERD and UNER English (CC BY-SA 4.0) and on entity labels written by Qwen2.5-7B-Instruct-AWQ (Apache-2.0) for FineWeb and Wikipedia passages, after a precision filter. The model therefore has share-alike lineage. CrossNER (MIT) only measures. Results, hand-check numbers and every hash are in recipes/openner/PROVENANCE.md and results.md.

The model is English only and well below the best published open-type recognizers on zero-shot sets (CrossNER 21 to 41 F1 on our run; recall is the weak side). When the types are fixed and you have labels, train the ordinary EntityRecognizer.

A saved head names its encoder, as every head does; load it with the encoder's directory available.

Published comparison

Strict token-span-and-type micro F1, threshold 0.5, frozen student-v1, shipped human + Qwen head. Intervals are sentence-bootstrap percentile 95% CIs (1,000 resamples, seed 1). All seven packed-model scores and intervals reproduce recipes/openner/results.md to the reported precision; neither weights nor thresholds were changed.

The consumer tools/OneNlp.OpenNerGate links the recipe's scorer but loads MultilingualModel.LoadOpenEntities() from packages. The measured combination is Models.Multilingual and FrozenEncoder 0.3.0-preview.1, core and Neural 0.4.0-preview.1. The newer scratch feed contained no multilingual package; the previously packed model was used, not repacked. Its embedded head (2,385,609 bytes, SHA-256 25d7cd3f8f3cc67b73aaca5d49db00dc88dc1758ffb34f43658aeb2fefff1d68), frozen encoder (aee56231d97134e455fe80a79738eb294c7974c61311412256fa9cd854eeeae6) and tokenizer (197d4cc5406ee12cc50c8b5511f2393cc32d9db321545979ce041c1199178356) were checked against the current manifests before inference. A newly packed multilingual 0.4.0-preview.1 is not verified by this run.

Models and reference

Reference: Zaratiana et al., GLiNER: Generalist Model for Named Entity Recognition using Bidirectional Transformer, arXiv:2311.08526v1, 14 November 2023, Table 1, §4.1 (zero-shot OOD benchmark); encoder and training in §3.2, evaluation in §3.4, decoding in §2.3. The paper is CC BY 4.0; PDF SHA-256 2b6efdc8337d12814308e480378587059a34a960d1da816f192b8cc0aaea04d0. Only the paper was read; no GLiNER implementation or weights were used.

These are the paper's counts, not measurements of downloaded GLiNER weights. Table 1 and §3.2 identify the encoder family and size variants, not exact checkpoint revisions. Their encoders are trained on Pile-NER; ours is frozen, and our head trains on Few-NERD, UNER and filtered Qwen labels. This is a comparison to other small open-type models, not a controlled architecture study.

Rows

“Position” compares the published S point estimate with our CI: above our upper bound is behind, below our lower bound is ahead, otherwise within noise. It is descriptive, not a paired significance test. Published GLiNER numbers have no CI. “Not comparable” means protocol equivalence is not established; it is not a passing quality verdict.

All five CrossNER rows are numerically behind even the smallest published GLiNER model, and behind M and L as well. The gaps to S are respectively 20.23, 19.10, 34.89, 29.62 and 29.60 F1 points. There is no evidence here that the published-small-model quality target is met; the protocol caveats do not turn these losses into passes. UNER and Few-NERD lack a same-dataset published GLiNER comparison in this paper.

The fixed-type black-box reference remains spaCy 3.8.18, 62.32 [59.45, 65.20] on UNER English-EWT test, trained with the documented default configuration on the same UNER training data. Its point estimate is below our lower CI bound (ahead by the point-vs-CI convention), but the intervals overlap and no paired spaCy comparison was run here. It is not an open-type or zero-shot baseline.

Protocol audit

• Splits: we read the fetched CrossNER domain test files, never train on them. The paper §3.4 names the OOD benchmark comprising CrossNER and MIT; it does not give file hashes, sentence counts or exact split revisions. Identical input rows cannot be certified from the paper alone.
• Filtering: the unchanged recipe excludes empty sentences, sentences over 120 tokens and tokens containing whitespace. It scores gold tokens, joined by single spaces, not a raw-text tokenization pipeline. Counts in the table are after filtering: all fetched UNER and CrossNER sentences remain; Few-NERD retains 37,637 of 37,648 (11 excluded). The paper does not specify these filters or enough input fingerprints to certify the same CrossNER inputs.
• Labels: our scorer asks for the union of types in the retained file on every sentence, including absent types as negatives, ordered ordinally. tools/openner/fetch.py expands labels to words, e.g. politicalparty → political party, programlang → programming language, otherperson → person, organization → organisation. The paper does not enumerate the exact strings or inventories requested for each CrossNER row, so label equivalence is unverified. Our type counts are politics 9, science 17, music 13, literature 12, AI 14; UNER 3; Few-NERD 67 derived labels (not the 66 original fine labels). “Zero-shot” here means no CrossNER examples used to train or select the head, not that every label word was unseen in training.
• Threshold and decoding: paper §2.3 keeps scores >0.5, with greedy flat non-overlap; ours keeps ≥0.5, also greedy flat non-overlap, up to 12 tokens. These have the same nominal threshold, but the exact-boundary convention differs. No threshold was retuned on test.
• Metric: paper §3.4 says exact-match entity F1. Ours is micro F1 over exact token range and type, sentence-bootstrapped, not relaxed matching. The paper supplies no bootstrap intervals or precise aggregation details.
• Few-NERD: our converter adds both fine and coarse labels for a mention where they differ, while prediction is flat and non-overlapping. Thus this seen-types row is the recipe's own two-level target, not a conventional fine-only Few-NERD benchmark score. It must not be compared to a published fine-only score without a separate, matching evaluation.

See the results table for the same rows, and tools/OneNlp.OpenNerGate/README.md for the command. This CPU-only measurement uses Linux x64, four threads; no training, GPU inference or speed claim is made.

Measurement environment and package fingerprints

Measured 8 November 2026 on Linux x64, AMD Ryzen 9 7950X3D (16 physical cores, 32 logical; AVX2, AVX-512F and FMA available), four reserved execution threads. .NET SDK 10.0.403, runtime 10.0.14, OneModel and OneModel.Training 1.0.0. No GPU was used. Load averages (1/5/15 min) were 2.97/3.72/7.08 before and 7.58/5.77/7.11 after; during-run load was not sampled. These are quality results, not timings. This measurement does not add a managed speed row: no managed open-type comparator is available.

Package SHA-256 fingerprints of the inputs to restore:

Data fingerprints are in recipes/openner/PROVENANCE.md and the fetched MANIFEST.tsv. The loaded dependency graph was checked to use the core, Neural and OneModel versions above, rather than the older transitive runtime.