atomizer-gliner2.5-base

atomizer on Fastino's GLiNER2.5 base (fastino/gliner2.5-base-v1) with atomizer's standard 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). 202M parameters.

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 (this repo)	Fastino GLiNER2.5 base	standard, no BiLSTM
atomizer-gliner2.5-base-native	Fastino GLiNER2.5 base	GLiNER2.5's own pretrained heads
atomizer-gliner2.5-base-xhead	Fastino GLiNER2.5 base	head-interaction block

Status: a research ablation. This repo holds the plain-heads baseline of the Fastino line: one version, v0.1 (= main). The native and xhead repos build on it; the model recommended for use is atomizer-gliner-small.

Model

The encoder of Fastino's GLiNER2.5 base (DeBERTa-v3-base, 12 layers; Apache-2.0) goes straight into atomizer's four heads, with no BiLSTM and no span layer (GLiNER2.5 has neither; GLiNER v1 backbones bring theirs pretrained):

Head	Predicts
anchor	per word: does an atom start here
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)
critic	per atom: verifiable, checkworthy

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

Load a tag with Atomizer.load("nishparadox/atomizer-gliner2.5-base", revision="<tag>").

Tag	Training	W&B run
v0.1	one run from fastino/gliner2.5-base-v1 on the v0.5 data (Propositionizer, ClaimBuster, CheckThat! 2024, synthetic v0.3, Decontextualization, Factcheck-Bench, chat templates); 3 epochs, 9,147 steps, batch 32, 1× A100; the last step	gliner25-base-nolstm

Until 2026-10-15 it was tag v0.6-rc2 on atomizer-gliner-small's dev branch; the files are the same (model.pt sha256 fbbe075f…).

A100 run of the v0.5 test suite (test splits only; soft F1 unless stated; resolved / wrong on Decontextualization), with the small release and the native model for comparison:

Model	synthetic	Factcheck-Bench	Decontext.	resolved	wrong ↓	Propositionizer	PropSegmEnt	CheckThat! F1 / AUC
v0.1 (this repo)	0.912	0.758	0.579	0.741	0.076	0.738	0.757	0.872 / 0.979
small v0.5 (163M)	0.905	0.753	0.551	0.703	0.073	0.742	0.757	0.865 / 0.978
native v0.1	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): this model − small v0.5: synthetic exact +0.032, soft +0.006, Decontextualization +0.027, no significant loss. Native v0.1 − this model: synthetic soft +0.007, Decontextualization +0.018, PropSegmEnt +0.010, no loss, which is why the native heads were kept and this line was not trained on the v0.7 data. The same encoder with a randomly initialised BiLSTM (W&B gliner25-base-full) tied with this model and was not published. All numbers: EVALS.md; the runs: EXPERIMENTS.md, runs 8 and 9.

Usage

The loading code is the atomizer package at NISH1001/atomizer (private for now, access on request). This model needs code ≥ v0.1.1; 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", revision="v0.1")

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))

Limitations

Those of the small model (its card), and ~1.6× slower than it on a CPU (12 encoder layers to its 6). It keeps first person as written ("I am Paradox. I won …" keeps "I").

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 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; atomizer-chat-synthetic was written by Claude).