atomizer-gliner2.5-base-native

atomizer on Fastino's GLiNER2.5 base (fastino/gliner2.5-base-v1) with its native heads: a single-pass span atomizer with a critic. It splits a text into short, self-contained claims ("atoms") and scores each one for verifiable (is it a checkable factual claim) and checkworthy (is it worth checking). 192M parameters, ~10 ms per text on a GPU, 60–70 ms on a laptop CPU.

The model does not generate text. It selects words from the input, replaces references ("She" → "Marie Curie") with earlier spans or keeps them as written, and adds closed-class glue words ("is", "the", "'s"). An atom cannot contain a name or number that is not in the input.

One repo per architecture. The atomizer models live in one Hugging Face repo per backbone and head design, so every tag of a repo loads the same way and has the same size and speed. Each repo has its own version numbers:

Repo	Backbone	Heads
atomizer-gliner-small	GLiNER small v2.5 (163M)	standard; the releases (main)
atomizer-gliner-large	GLiNER large v2.5	standard
atomizer-gliner2.5-base	Fastino GLiNER2.5 base	standard, no BiLSTM
atomizer-gliner2.5-base-native (this repo)	Fastino GLiNER2.5 base	GLiNER2.5's own pretrained heads
atomizer-gliner2.5-base-xhead	Fastino GLiNER2.5 base	head-interaction block

Status: research models. main is the latest version (v0.2); every version is a tag. The model recommended for use is atomizer-gliner-small v0.6: trained on the same data, within noise of this one, and lighter.

Model

The encoder of Fastino's GLiNER2.5 base (DeBERTa-v3-base, 12 layers; Apache-2.0) goes straight into four heads, two of them from GLiNER2.5 itself:

Head	Predicts	From GLiNER2.5
anchor	per word: does an atom start here	matched against an [E] claim query, initialised from GLiNER2.5's own encoding of that schema token
pair	per (atom, word): KEEP / SUB / DROP, and glue after the word	–
antecedent	per reference: which earlier span it means, or none (keep the reference as written)	GLiNER2.5's pretrained relation scorer (a port of the gliner2 package's SparseRelationScorer), with a refers_to query
critic	per atom: verifiable, checkworthy	atom words matched against [L] verifiable / [L] checkworthy queries

GLiNER2.5 reads its queries from a schema prompt in the input; atomizer's task never changes, so the queries are computed once from GLiNER2.5's encoding of the schema and then learned, and texts go in without a prompt. Assembly from the predictions is deterministic; texts longer than 1,024 tokens are atomized in sentence windows with read-only context. Details: ARCHITECTURE.md.

Versions

Each tag is one run from fastino/gliner2.5-base-v1 on all the training data (3 epochs, batch 32, 1× A100; the last step is published). Load a tag with Atomizer.load("nishparadox/atomizer-gliner2.5-base-native", revision="<tag>").

Tag	Training data	W&B run	Notes
v0.2	the v0.6 data (the same as small v0.6): Propositionizer, ClaimBuster, CheckThat! 2024, synthetic v0.4, Decontextualization, Factcheck-Bench, WiCE, FActScore, chat templates; 9,690 steps	native-v07	the fewest wrong references of that round's candidates (5.8%)
v0.1	the v0.5 data: the same without WiCE, FActScore and synthetic v0.4; 9,147 steps	gliner25-native	the best model on the old test suite (overall 0.788)

Until 2026-10-15 these two were tags v0.7-rc2 and v0.6-rc3 on atomizer-gliner-small's dev branch; the files are the same (model.pt sha256 6d5889e6… and f7fe213e…).

A100 runs of the test suite (test splits only; soft F1 unless stated; resolved / wrong on Decontextualization; the v0.7 suite adds WiCE and FActScore and a larger synthetic test, so the two suites are not comparable with each other):

Tag	suite	synthetic	Factcheck-Bench	Decontext.	resolved	wrong ↓	Propositionizer	PropSegmEnt	CheckThat! F1 / AUC	WiCE soft / exact	FActScore soft / exact
v0.2	v0.7	0.899	0.766	0.577	0.728	0.058	0.738	0.779	0.856 / 0.977	0.813 / 0.318	0.706 / 0.084
v0.1	v0.7	0.906	0.758	0.597	0.728	0.067	0.737	0.766	0.867 / 0.976	0.795 / 0.232	0.603 / 0.014
v0.1	v0.5	0.918	0.758	0.597	0.728	0.067	0.737	0.766	0.867 / 0.976	–	–

Paired bootstraps (2,000 resamples over texts): v0.2 − v0.1 wins on FActScore +0.102, WiCE exact +0.087 / soft +0.017 and PropSegmEnt +0.012, loses on Decontextualization −0.020 and synthetic soft −0.007. v0.2 vs the small v0.6 (163M, the same data): no significant difference, so the small model is the proposed release. All numbers and the other pairs: EVALS.md; the runs: EXPERIMENTS.md, runs 10 and 12.

Usage

The loading code is the atomizer package at NISH1001/atomizer (private for now, access on request). This architecture needs code ≥ v0.2.0; v0.3.0 is current.

uv add "atomizer @ git+ssh://git@github.com/NISH1001/atomizer@v0.3.0"

from atomizer import Atomizer

atomizer = Atomizer.load("nishparadox/atomizer-gliner2.5-base-native", revision="v0.2")

for atom in atomizer.atomize("Tesla was founded in 2003. It delivered about 1.8 million vehicles in 2023."):
    print(atom.text, round(atom.verifiable, 2), round(atom.checkworthy, 2))

atomizer.atomize(text, min_verifiable=0.5)      # only checkable facts
print(atomizer.explain(text).show())            # every head's predictions

Limitations

Those of the small model (its card), plus: ~1.6× slower than the small model on a CPU (12 encoder layers to its 6); v0.1 collapses on heavily repetitive text (100+ near-identical sentences); v0.2 resolves "I" and "my" to a self-given name, but writes the possessive ("Paradox's lab") less reliably than the small v0.6.

License

The weights are CC-BY-4.0; the atomizer code is MIT. Fine-tuned from Fastino's gliner2.5-base-v1 (Apache-2.0); the relation scorer is a port of the gliner2 package's SparseRelationScorer (Apache-2.0). The training data has its own terms (Propositionizer labels were written by GPT-4; ClaimBuster is CC-BY-SA 4.0; CheckThat! 2024 under the CLEF lab's terms; WiCE ODC-BY; FActScore's data terms are not stated and it is not redistributed; atomizer-chat-synthetic was written by Claude).