moderation[border]

The moderation detector for border, an embeddable library that inspects the text going into and coming out of an LLM and returns a structured decision plus an audit-grade evidence record.

flowxai/moderation on the hub. It is one detector of 28, and it is not a general purpose moderation classifier: it was trained for this library's policy, is read at the operating point below, and reports through the evidence record rather than returning a bare score.

This card is generated from the evaluation and export artifacts of the training run, so every number on it is reproducible from this repository rather than asserted.

What it is

• Base model: FacebookAI/xlm-roberta-base
• Head: multi_label_classification
• Labels: cyber_intrusion, defamation, election_integrity, extremism, fraud_deception, hate_incitement, illicit_drugs, property_crime, self_harm, sexual_exploitation, violent_facilitation, weapons_cbrn
• Artifact: onnx/model.int8.onnx, 535 MB, opset 17
• Trained at: 96 tokens

12 labels, and what is missing

The taxonomy this detector implements defines 13 labels. This head was trained on 12. The rest are listed here rather than omitted, because a card that names only what a model can do presents partial coverage as complete.

• child_safety: must not be synthetically generated. The label covers sexualisation of minors and grooming, and no length band or framing makes generating either acceptable. It needs a vetted source with its provenance recorded, which is a decision about provenance rather than compute. Until then moderation is a twelve-label head and the thirteenth reports unavailable rather than clean.

In flowx-border a label with no training behind it is reported as unavailable rather than as clean. A detector that silently scores zero on something it cannot see is indistinguishable from one that looked and found nothing.

Operating point

Threshold 0.84, calibrated on the validation split against the macro_f1 objective.

This number is not decoration. Read at the 0.5 default that looked reasonable, several detectors in this family reported F1 0.000 in every language, because their scores separate positives from negatives well below 0.5. One of them went from 0.000 to 0.893 on the threshold alone. Use the value above, or calibrate your own on your own data.

• At the 0.5 default: 0.923
• At the calibrated 0.84: 0.931

How to use it

Through the library, which is what this model is for. It loads the artifact below, applies the operating point above, and returns a decision with an evidence record rather than a bare score.

pip install flowx-border

# policy.yaml
policy_id: default
version: 1

detectors:
  moderation:
    enabled: true
    on_fail: flag
    threshold: 0.84

from flowx_border import load_policy, scan_input, scan_output

policy = load_policy("policy.yaml")

decision = scan_input(user_text, policy)
decision = scan_output(model_answer, policy)

print(decision.verdict)      # allow | flag | redact | block
print([f.label for f in decision.findings if f.detector_id == "moderation"])
print(decision.evidence.record_id)

This detector reads the input and output side, so scan_input and scan_output is where it fires. It is T2, so it runs on the standard path and can be disabled per policy. Its budget is 150 ms at 87 tokens on one CPU thread.

The weights are fetched once and cached, and a scan needs no network after that. Nothing here calls out to a hosted model, and the evidence record carries hashes rather than your text.

Without the library

The artifact is plain ONNX, so it will load in onnxruntime directly. Two things you then own yourself, and they are the reason the library exists: the operating point above is not in the graph, and neither is the chunking. Inputs longer than the trained window have to be split and recombined, or the scores past it are extrapolation.

import onnxruntime as ort
from huggingface_hub import hf_hub_download
from tokenizers import Tokenizer

repo = "flowxai/moderation"
session = ort.InferenceSession(hf_hub_download(repo, "onnx/model.int8.onnx"))
tokenizer = Tokenizer.from_file(hf_hub_download(repo, "tokenizer.json"))

Per language

Per language rather than an aggregate, because an aggregate across 26 languages hides the tail and the tail is the point.

Language	Support	P	R	F1	Note
cs Czech	60	1.000	1.000	1.000
el Greek	60	1.000	1.000	1.000
et Estonian	60	1.000	1.000	1.000
fi Finnish	60	1.000	1.000	1.000
fr French	59	1.000	1.000	1.000
hu Hungarian	60	1.000	1.000	1.000
it Italian	59	1.000	1.000	1.000
pl Polish	60	1.000	1.000	1.000
ro Romanian	60	1.000	1.000	1.000
sl Slovenian	59	1.000	1.000	1.000
sv Swedish	60	1.000	1.000	1.000
lt Lithuanian	60	0.984	1.000	0.992
tr Turkish	60	0.984	1.000	0.992
da Danish	60	1.000	0.983	0.992
de German	60	1.000	0.983	0.992
en English	59	0.983	1.000	0.992
nl Dutch	60	1.000	0.983	0.992
sk Slovak	60	1.000	0.983	0.992
es Spanish	59	1.000	0.983	0.992
hr Croatian	58	0.983	1.000	0.992
pt Portuguese	60	0.968	1.000	0.984
az Azerbaijani	60	1.000	0.967	0.983
bg Bulgarian	60	1.000	0.967	0.983
lv Latvian	60	1.000	0.967	0.983
ga Irish	60	0.952	0.983	0.967
mt Maltese	59	0.983	0.949	0.966	not in base model pretraining

Weakest languages

Published rather than dropped. A coverage table with the bad rows removed is not a coverage table.

• mt Maltese: F1 0.966 (absent from XLM-R pretraining, which is a base-model limit)
• ga Irish: F1 0.967
• az Azerbaijani: F1 0.983

Quantisation

The published artifact is INT8, and only the embedding table is quantised.

Quantising everything is what most examples do and it does not work for this base model. Measured on 300 real test texts at the detector's own threshold:

Recipe	Size	Mean logit drift	Decisions changed
all ops (the usual default)	279 MB	0.68	51 / 300
MatMul only	856 MB	0.64	48 / 300
Gather only, what ships here	535 MB	0.0036	0 / 300

The embedding table carries the whole size win at no accuracy cost, while quantising the encoder MatMuls changes one decision in six to save 256 MB. XLM-RoBERTa has large activation outliers and per-tensor dynamic quantisation of activations is exactly what they defeat.

For this artifact specifically: 1 of 300 decisions differ from the fp32 checkpoint, mean logit drift 0.0028, read as sigmoid_at_threshold. A quantised model that answers differently is a different detector, so this is measured rather than assumed.

Limitations

• Synthetic training data. Generated natively per language, never translated from English, so the sentence structure is the target language's own. It is still synthetic, and a production distribution will differ.
• Maltese is absent from XLM-RoBERTa's pretraining set. That is a fact about the base model, and it is not an explanation for a weak score. This card said "no amount of data fixes that" until 2026-08-14, which this project's own measurement disproves: the nsfw detector scored 0.000 in Maltese, was blamed on the base model, and went to 1.000 with perfect precision and recall when its corpus went from 2 positives per language to 10. Nothing about the model changed. So where a language scores badly here, read the support column first.
• This is not a compliance product. It produces evidence about controls that were applied. It does not make anyone compliant with anything, and the obligations under the EU AI Act sit with the provider or deployer of a system, not with a model or a library.

Licence

Apache-2.0, declared in the metadata above as well as here, so that a tool reading the repository can attest it rather than a human having to read prose.