Model Card for Model ID

This model is a facebook/bart-large fine-tuned on non-toxic inputs from lmsys/toxic-chat dataset.

Model Details

This model is not intended to be used for plain inference despite it is unlikely to generate toxic content. It is intended to be used instead as "utility model" for detecting and fixing toxic content as its token probability distributions will likely differ from comparable models not trained/fine-tuned over non-toxic data.

Its name tci_plus refers to the G+ model in Detoxifying Text with MaRCo: Controllable Revision with Experts and Anti-Experts.

It can be used within TrustyAI's TMaRCo tool for detoxifying text, see https://github.com/trustyai-explainability/trustyai-detoxify/.

Model Description

Developed by: [tteofili]
Shared by: [tteofili]
License: [AL2.0]
Finetuned from model: ["facebook/bart-large"]

Uses

This model is intended to be used as "utility model" for detecting and fixing toxic content as its token probability distributions will likely differ from comparable models not trained/fine-tuned over toxic data.

Bias, Risks, and Limitations

This model is fine-tuned over non-toxic inputs from the lmsys/toxic-chat dataset and it is very likely to produce toxic content. For this reason this model should only be used in combination with other models for the sake of detecting / fixing toxic content.

How to Get Started with the Model

Use the code below to start using the model for text detoxification.

from trustyai.detoxify import TMaRCo
tmarco = TMaRCo(expert_weights=[-1, 3])
tmarco.load_models(["trustyai/tci_minus", "trustyai/tci_plus"])
tmarco.rephrase(["white men can't jump"])

Training Details

This model has been trained on non-toxic inputs from the lmsys/toxic-chat dataset.

Training Data

Training data from the lmsys/toxic-chat dataset.

Training Procedure

This model has been fine tuned with the following code:

from trustyai.detoxify import TMaRCo

dataset_name = 'lmsys/toxic-chat'
data_dir = ''
perc = 100
td_columns = ['model_output', 'user_input', 'human_annotation', 'conv_id', 'jailbreaking', 'openai_moderation',
              'toxicity']

target_feature = 'toxicity'
content_feature = 'user_input'
model_prefix = 'toxic_chat_input_'
tmarco.train_models(perc=perc, dataset_name=dataset_name, expert_feature=target_feature, model_prefix=model_prefix,
                    data_dir=data_dir, content_feature=content_feature, td_columns=td_columns)

Training Hyperparameters

This model has been trained with the following hyperparams:

training_args = TrainingArguments(
    evaluation_strategy="epoch",
    learning_rate=2e-5,
    weight_decay=0.01
)

Evaluation

Testing Data, Factors & Metrics

Testing Data

Test data from the lmsys/toxic-chat dataset.

Metrics

The model was evaluated using perplexity metric.

Results

Perplexity: 1.04

trustyai
/

tci_plus