README.md

metadata

license: mit
datasets:
  - akaruineko/offensively-neutral
language:
  - en
base_model:
  - distilbert/distilbert-base-uncased
pipeline_tag: text-classification
library_name: transformers
tags:
  - moderation
  - filtering
  - offensive
  - clean
model-index:
  - name: akaruineko/ftan-2.0
    results:
      - task:
          type: text-classification
          id: overall
        dataset:
          type: akaruineko/ftanch
          name: FTANch
          split: test
        metrics:
          - type: acc
            value: 0.8144
            name: Accuracy
          - type: p
            value: 0.7637
            name: Precision
          - type: r
            value: 0.989
            name: Recall
          - type: f1
            value: 0.8619
            name: F1-Score
      - task:
          type: text-classification
          id: test
        dataset:
          type: akaruineko/ftanch
          name: FTANch
          split: test
        metrics:
          - type: acc
            value: 0.7752
            name: Accuracy
          - type: p
            value: 0.6938
            name: Precision
          - type: r
            value: 0.9852
            name: Recall
          - type: f1
            value: 0.8142
            name: F1-Score
      - task:
          type: text-classification
          id: test_obfuscated
        dataset:
          type: akaruineko/ftanch
          name: FTANch
          split: test
        metrics:
          - type: acc
            value: 0.8979
            name: Accuracy
          - type: p
            value: 0.8866
            name: Precision
          - type: r
            value: 0.9942
            name: Recall
          - type: f1
            value: 0.9373
            name: F1-Score
      - task:
          type: text-classification
          id: plain
        dataset:
          type: akaruineko/ftanch
          name: FTANch
          split: test
        metrics:
          - type: acc
            value: 0.667
            name: Accuracy
          - type: p
            value: 0.4398
            name: Precision
          - type: r
            value: 0.9622
            name: Recall
          - type: f1
            value: 0.6037
            name: F1-Score
      - task:
          type: text-classification
          id: mutated
        dataset:
          type: akaruineko/ftanch
          name: FTANch
          split: test
        metrics:
          - type: acc
            value: 0.8979
            name: Accuracy
          - type: p
            value: 0.8866
            name: Precision
          - type: r
            value: 0.9942
            name: Recall
          - type: f1
            value: 0.9373
            name: F1-Score

ftan-2.0

ftan-2.0 is a fine-tuned DistilBERT sequence classification model for detecting offensive text.

The model predicts one of two labels:

• clean — non-offensive text
• offensive — offensive text

Training

ftan-2.0 is the continuation of the akaruineko/bad-good-classifier-ru_en project.

The new version was trained on the akaruineko/offensively-neutral dataset, containing approximately 1.3 million text samples.

Training used a larger dataset than the previous model and included evaluation across multiple epochs to select the best-performing checkpoint.

The best checkpoint was selected based on evaluation performance rather than simply using the final training checkpoint.

Usage

from transformers import pipeline

classifier = pipeline(
    "text-classification",
    model="akaruineko/ftan-2.0"
)

result = classifier("you are stupid")
print(result)

Example:

[{'label': 'offensive', 'score': 0.965}]

Intended Use

ftan-2.0 can be used for:

• content moderation
• filtering offensive messages
• dataset preprocessing
• text classification experiments
• moderation pipelines
• research and experimentation with text classifiers

Limitations

This model should not be treated as a perfect moderation system.

Offensiveness can depend heavily on context, intent, quotation, sarcasm, reclaimed language, and the surrounding conversation. The model may therefore produce incorrect predictions for ambiguous or context-dependent text.

For example, a sentence discussing an offensive word may still receive a non-trivial offensive score even when the sentence itself is not an insult.

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

Example Predictions

Some example inference results:

"b****" # censored
→ offensive (0.988)

"you are stupid"
→ offensive (0.965)

"the word \"stupid\" is offensive"
→ offensive (0.695)

"beach"
→ clean (0.922)

These examples are illustrative and should not be interpreted as a formal benchmark.

License

See the repository/model files for the applicable license.

Author

Created by akaruineko.

This model is the 2.0 continuation of the bad-good-classifier-ru_en project.