Comparison · ReDefiNER
ReDefiNER vs GLiNER
Updated 19 November 2026, by Jungroo AI Labs, the makers of ReDefiNER
Both are open models that find the entity types you choose when you run them, with no retraining, and both run on a CPU. GLiNER matches text against type names. ReDefiNER reads a rulebook: a one-line description and optional rules for each type. On the same 7 test sets, ReDefiNER scored 63.1 F1 with descriptions against GLiNER's 61.7. With type names alone, GLiNER scored 61.6 against ReDefiNER's 56.2. We make ReDefiNER, and we ran both models ourselves.
The short answer
- Choose GLiNER when your types are clear from their names (person, company, city), when you want a smaller and widely used model, or when you don't want to write descriptions.
- Choose ReDefiNER when a type needs explaining, or needs rules such as “leave out titles”, “never label our own brand” or “a reference starting with # is an order number”, and when you want to change what counts by editing text.
- Test both on 20 to 50 of your own sentences. On a single dataset the gap can be much larger than these averages, in either direction.
Accuracy on the same test
Mean F1 over the 7 CrossNER and MIT test sets (AI, literature, music, politics, science, movies and restaurants; 6,470 sentences). Each type was given either as a one-line description or by its name alone. For GLiNER and NuNER, the first number uses the model's default threshold and the second a threshold chosen on the dev sets. ReDefiNER uses its one fixed release setting.
| Model | With descriptions | Names only |
|---|---|---|
| ReDefiNER 0.8B | 63.1 | 56.2 |
| GLiNER large v2.5, our run | 61.7 / 62.2 | 61.6 / 63.1 |
| NuNER Zero-span 0.4B, our run | 26.5 / 55.5 | 62.6 / 62.6 |
| GNER-T5-large 0.78B, as reported | – | 63.5 |
F1 counts a prediction as right only when both the span and the type match exactly. GLiNER and NuNER were run through the gliner library 0.2.29 on the same sentences, with each type's description, or its name, as the label. Full tables, settings and rulebooks.
In short: with descriptions, ReDefiNER is ahead of GLiNER at either threshold. GLiNER's best setting (names only, threshold tuned on dev data) ties ReDefiNER's best at 63.1. With names alone, ReDefiNER is the weakest model in this table.
How they differ
| GLiNER large v2.5 | ReDefiNER | |
|---|---|---|
| What you give it | A list of type names (labels) | A rulebook: types, a one-line description each, numbered rules and optional examples |
| Rules such as “leave out titles” | No place to state them, beyond the label text | Written in plain English. Which rules work |
| How it works | A bidirectional encoder that matches text spans against the labels | A 0.8B language-model backbone with span and type heads. Non-generative: it never writes text |
| What it returns | Spans with a label and a score | Exact character spans, each with a type and a confidence |
| Size | A smaller encoder model | 0.8 billion parameters, one 0.77 GB int8 file |
| Licence | Apache 2.0 (per its model card) | Apache 2.0 |
Where ReDefiNER is weaker
- Type names alone. 56.2 against GLiNER's 61.6. Always write a one-line description for each type.
- Abstaining. When a type's condition isn't clearly met, it tends to label anyway.
- Rules against a strong prior are mostly ignored, such as relabelling a person's name as something other than a person, or “only revenue amounts, not every money amount”.
- Limits. English only, and spans of up to 16 tokens in release 1.0.8.
- Speed. On a 4-core CPU server, each sentence takes about 0.4 seconds once the rulebook is loaded; reading a new rulebook of about 1,000 tokens takes 4 to 5 seconds first. We have not timed GLiNER on the same machine.
Try both on your own text
Both install with pip and run on a CPU. Give GLiNER your type names, and give ReDefiNER the same types with a one-line description each:
pip install gliner redefiner
from gliner import GLiNER
gliner = GLiNER.from_pretrained("gliner-community/gliner_large-v2.5")
gliner.predict_entities(text, ["person", "organization"], threshold=0.5)
from redefiner import Tagger
tagger = Tagger.from_pretrained("jungroo-ai/redefiner-0.8b")
tagger.extract(text, """[RULES]
Extract every entity of the supplied types. Return exact spans from the text.
[TYPE] PERSON
[DESCRIPTION] A person's name.
PE1 Label the full name; leave out titles such as "Dr" or "Prof".
[TYPE] ORGANIZATION
[DESCRIPTION] A company, university or other institution.
""")