t5-base fine-tuned for text detoxification (ParaDetox)

Sequence-to-sequence text style transfer: rewrites English toxic/offensive text into neutral text while preserving the propositional content. Fine-tuned from google-t5/t5-base on s-nlp/paradetox.

Task formulation

Single-direction conditional generation with an explicit T5 task prefix:

input : detoxify: <toxic sentence>
target: <neutral sentence>

The prefix detoxify: is applied identically during training, evaluation and inference.

Model

Training data

s-nlp/paradetox ships a single train split of parallel toxic/neutral sentence pairs. Validation and test splits were carved out of it.

Because a toxic sentence can carry up to three different crowd-written neutral paraphrases, the split is computed over unique source sentences, not over rows. A row-level random split would place two paraphrases of the same sentence in train and test and inflate every test metric.

Split strategy: group-aware over unique source sentences, seed 42. Split fingerprints (sha256 over the sorted unique sources, first 16 hex chars): train=5107f460c600b24a, validation=1553707e06c53e22, test=110b37a06c3c22fb.

Training hyperparameters

Checkpoint selection: best validation loss (eval_loss). The test split was not used for model or hyperparameter selection.

Best validation loss: 0.7999.

Generation configuration

Deterministic beam search; identical settings were used for every number below.

Evaluation

Held-out test split: 596 unique toxic inputs, 962 human references (multi-reference scoring).

1. Reference-based metrics

Measures overlap with human detoxifications. Necessary but not sufficient: a correct detoxification worded differently from the reference is penalised here. The copy-input column scores the unchanged toxic input against the same references and shows how much of any BLEU/ROUGE is simply rewarded for copying.

2. Classifier-based toxicity

Classifier: s-nlp/roberta_toxicity_classifier, threshold 0.5.

Detoxification rate = among inputs the classifier calls toxic, the share whose generated output the classifier calls non-toxic.

The classifier is a noisy proxy, not ground truth: false positives on quoted speech and reclaimed slurs, false negatives on implicit toxicity. The human references themselves do not reach a 0% toxicity rate, which bounds how much of the remaining rate is really model error.

3. Semantic preservation

Embeddings: sentence-transformers/all-MiniLM-L6-v2.

cos(input, human reference) is the practical ceiling: human annotators also drop content when they remove toxic spans.

4. Joint diagnostic

Among the 596 inputs classified toxic, 83.56% produced an output that is both classified non-toxic and stays within cosine 0.6 of the input.

This is an intersection of two noisy measurements, reported as a diagnostic. It is not an overall model score.

Example outputs

Seeded random sample of the test split, not selected for quality.

Usage

from transformers import AutoTokenizer, AutoModelForSeq2SeqLM

tokenizer = AutoTokenizer.from_pretrained("thealper2/t5-base-paradetox-detox")
model = AutoModelForSeq2SeqLM.from_pretrained("thealper2/t5-base-paradetox-detox")

text = "You are an idiot and nobody likes you."
inputs = tokenizer("detoxify: " + text, return_tensors="pt")
outputs = model.generate(
    **inputs,
    num_beams=5,
    max_new_tokens=64,
    length_penalty=1.0,
    no_repeat_ngram_size=3,
    early_stopping=True,
    do_sample=False,
)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

The task prefix detoxify: is required. Without it the model is out of distribution.

Limitations

• English only. ParaDetox is an English corpus of Reddit/Twitter/Jigsaw-style comments.
• Trained on short, single-sentence inputs (source p99 = 38 tokens). Long or multi-sentence input is truncated at 64 tokens.
• The dominant training pattern is toxic-span removal. The model deletes or softens offensive spans; it does not reliably rewrite implicit toxicity, sarcasm, dog whistles or threats phrased politely.
• Detoxification and semantic preservation trade off against each other. Deleting a whole clause maximises the toxicity metric and damages the similarity metric; the two tables above must be read together.
• Not a content-moderation system. It rewrites text, it does not detect, filter or classify it, and it must not be used as a safety filter.
• No number here measures fluency or grammaticality.
• Test metrics are computed with a classifier and an embedding model that have their own biases, including known demographic biases in toxicity classifiers.

Reproduction

python -m src.train
python -m src.evaluate
python -m src.qualitative

Config, dataset split fingerprints and the full training record are stored next to the weights (experiment_config.yaml).