Experimental model

FTAN-3.0-multi is an experimental release.

This version introduces multi-label moderation and is still under active development. Its performance and behavior may differ significantly from previous FTAN releases.

For general-purpose binary offensive-text classification, use akaruineko/ftan-2.5 instead.

FTAN-3.0-multi is primarily intended for experimentation with multi-label moderation, research, and development.

FTAN-3.0-multi

FTAN-3.0-multi is a multi-label text moderation model from the FTAN model family.

Unlike previous FTAN versions, which focused on binary offensive-text classification, FTAN-3.0-multi can assign multiple moderation labels to the same text.

This release represents a major step in the FTAN series, moving from binary offensive classification to a multi-label moderation taxonomy.

Model Details

Model: FTAN-3.0-multi
Architecture: Transformer-based sequence classifier
Task: Multi-label text classification
Base architecture: DistilBERT
Labels: 3
Training dataset: akaruineko/multi-awesome-moderation
Dataset size: approximately 1.84 million examples
Training approach: teacher-generated pseudo-labels / knowledge distillation
Teacher model: KoalaAI/Text-Moderation

Labels

FTAN-3.0-multi predicts three moderation categories:

Label	Meaning
`H`	Hate — content that expresses, incites, or promotes hate based on protected characteristics.
`HR`	Harassment — content that may be used to torment or annoy individuals in real life, or make harassment more likely to occur.
`H2`	Hate / threatening — hateful content that also includes violence or serious harm towards a targeted group.

Because this is a multi-label classifier, multiple labels may be predicted for a single text.

For example, a text may receive:

H  → 0.91
HR → 0.84
H2 → 0.12

The model therefore does not force every input into exactly one category.

Training

FTAN-3.0-multi is a major continuation of the FTAN series.

Earlier FTAN models primarily focused on binary offensive-text classification:

FTAN 1.x
    ↓
word-level classification

FTAN 2.x
    ↓
binary text classification
    ↓
clean / offensive

FTAN 3.0-multi
    ↓
multi-label moderation
    ↓
H / HR / H2

The model was trained on akaruineko/multi-awesome-moderation, containing approximately 1,844,543 text examples.

The dataset was created from the FTAN dataset lineage and processed using teacher-generated moderation predictions from KoalaAI/Text-Moderation.

The dataset construction pipeline applies filtering, confidence thresholds, duplicate removal, text normalization, and label validation before examples are included in the final training set.

Knowledge Distillation

The moderation labels were generated using KoalaAI/Text-Moderation, which acts as the teacher model.

The resulting predictions were then filtered and converted into the FTAN moderation taxonomy before being used to train FTAN-3.0-multi.

Conceptually:

FTAN dataset
     ↓
KoalaAI/Text-Moderation
     ↓
teacher predictions
     ↓
filtering + validation
     ↓
multi-awesome-moderation
     ↓
FTAN-3.0-multi

This makes the model a teacher-student / pseudo-labeling experiment, rather than a model trained exclusively on manually annotated labels.

Usage

from transformers import pipeline

classifier = pipeline(
    "text-classification",
    model="akaruineko/ftan-3.0-multi",
    top_k=None
)

result = classifier("you are stupid")

print(result)

Example output:

[
    [
        {'label': 'H', 'score': 0.68},
        {'label': 'H2', 'score': 0.00},
        {'label': 'HR', 'score': 0.00}
    ]
]

For multi-label classification, top_k=None can be used to obtain the scores for all labels.

Intended Use

FTAN-3.0-multi can be used for:

research and experimentation
teacher-student / knowledge-distillation experiments
dataset preprocessing
multi-label moderation research
content filtering experiments

Limitations

FTAN-3.0-multi should not be treated as a perfect moderation system.

The model was trained using automatically generated pseudo-labels. Therefore, it may reproduce errors, biases, or inconsistencies present in the teacher model and the dataset-generation pipeline.

Moderation decisions can also depend heavily on context, intent, quotation, sarcasm, reclaimed language, and surrounding conversation.

The model operates on individual text inputs and does not have access to conversation history unless it is explicitly provided as input.

The three-label taxonomy used by this model is also intentionally limited and does not represent every possible form of harmful or inappropriate content.

For high-stakes moderation, predictions should be evaluated using an independent, human-verified evaluation set.

Version History

FTAN-3.0-multi represents a significant change in the FTAN family.

FTAN 1.0
    ↓
word classification

FTAN 2.0
    ↓
binary text classification

FTAN 2.5
    ↓
incremental continuation of FTAN 2.x

FTAN 3.0-multi
    ↓
multi-label text moderation

The transition from FTAN 2.x to FTAN 3.0-multi changes the task from binary classification to multi-label moderation and introduces a new teacher-generated moderation dataset.