DeliVali/AIDAM_VREAL_279M_V0.1.0

AIDAM verifier — small comparative-logic NLI core

The verification core of AIDAM, an open fact-checking agent by Jeffrey Romero Del Val: a ~0.3B multilingual NLI model that judges (evidence, claim) pairs — supports / neutral / refutes — and an auditable aggregation layer that turns pair judgments into verdicts with citations. Verdicts always come from this core plus explicit math; LLMs assist retrieval and narration but never judge (measured: LLM-as-sole-judge 24.0% vs the aggregator's 58.0% on AVeriTeC-100).

Benchmarks (honest, all measured; details in the repo ROADMAP)

Benchmark	AIDAM best	Reference
LLM-AggreFact (BAcc, 11 subsets)	71.0 (v20, 0.3B)	MiniCheck-FT5 (0.77B): 74.7
FEVER (label acc., oracle evidence)	86.0 (v11 specialist)	large fine-tuned models: ~88-92
SciFact (label acc., oracle)	66.3 (v15)	—
AVeriTeC dev-500 (offline store)	62.6 (v10, production)	majority baseline: 61.0

Every promoted version passed pre-registered gates; rejected experiments are documented alongside the promotions (label-poisoning audits, register interference laws, backbone trade-offs) in docs/ROADMAP.md.

The agent around the core

The model ships inside a full local agent (GPU, CPU/ONNX, or 319 MB quantized): multi-source multilingual retrieval with a calibrated meaning-level evidence filter, question answering with cited sentences, clarification questions when evidence splits into distinct senses, conversational context (keyword-graph memory), sandboxed measured code comparison, native file control with permission cards, and an OpenAI-compatible endpoint so assistant gateways (OpenClaw-style) can use AIDAM from any messenger.

Training

mDeBERTa-v3 and mmBERT lineages fine-tuned on VitaminC + MNLI/ANLI plus purpose-built, shortcut-proofed pair sets (denial patterns, scientific register, FEVER register, D2C long-document and summary-shaped pairs, cross-claim abstention pairs) — each dataset promoted only after moving a benchmark behind pre-set bars. Recipes and generators are open in training/.

The AIDAM model family — one verifier per register

Model	Register it judges	Backbone	Headline number
AIDAM_VNEWS_308M_V0.1.0	news summaries, grounding	mmBERT 308M	LLM-AggreFact 71.0
AIDAM_VWIKI_279M_V0.1.0	encyclopedic claims	mDeBERTa-v3 279M	FEVER 86.0
AIDAM_VSCI_279M_V0.1.0	scientific writing	mDeBERTa-v3 279M	SciFact 66.3
AIDAM_VREAL_279M_V0.1.0	real-world and viral claims	mDeBERTa-v3 279M	AVeriTeC-500 62.6
AIDAM_CODING_1.5B_V0.1.0	(not a verifier — code expert)	Qwen2.5-Coder 1.5B	MBPP-test 63.0%

Why five verifiers and not one. A register is the dialect a text is written in: a news summary, a scientific abstract, a Wikipedia article and a viral post state facts in structurally different ways. Training one small verifier on a new register measurably costs it capacity in the others — 30k encyclopedic-register pairs pulled the model toward that style and its score on viral claims fell from 62.6 to 60.2. At 279M capacity the AVeriTeC-optimal and the grounding-optimal models had genuinely diverged into different checkpoints.

So each model above is the champion of one register and is beaten on the others by its own siblings. That is not a hedge; it is the measurement. In AIDAM they are selected per claim and several are consulted at once, because betting on a single routing choice would produce a confident wrong verdict.

A specialist's score counts only on the benchmark it was built for. Quoting the SciFact number as if it described general capability would be the same self-deception as moving a gate after seeing the result.

Versioning. AIDAM_<SPECIALISATION>_<PARAMETERS>_<VERSION>, and the parameter count is the real one. These models are V0.1.0: the research era, the line that produced the measurements above. V1.0.0 is reserved for the models the project is about to train, which do not exist yet — so a name never designates two artefacts at once. Earlier names still redirect here.

Formats

PyTorch (safetensors) and ONNX are both published. ONNX is what lets this model run on a CPU without a multi-gigabyte deep-learning runtime — roughly 50 MB of dependencies instead of ~3 GB — which is what makes it affordable to consult several verifiers about the same claim on a machine with a 4 GB GPU.

The ONNX export was accepted only after being checked against the PyTorch weights on real (evidence, claim) pairs: identical predicted labels, and a maximum logit difference on the order of 1e-6, which is floating-point noise rather than disagreement. Same weights, different runtime.

License and credit

Apache-2.0. Please keep visible credit to Jeffrey Romero Del Val (NOTICE file; citation in CITATION.cff). Information is free: the only arbiter is evidence.