metadata

license: mit
datasets:
  - OxAISH-AL-LLM/wiki_toxic
  - textdetox/multilingual_toxic_spans
language:
  - en
base_model:
  - openai-community/gpt2

Model Card for Toxic Text GEN

This model is a decision Tranformer for text generation with controlled toxicity (0-1).

Model Details

Model Description

Made using a decision transformer, it can generate toxic sentences based on a toxicity control (defined as reward-to-go/rtg).

Current text generation is not very coherent due to lack of variety in training data and low compute.

Developed by: [Ashed00]
Finetuned from model: [GPT-2]

Model Sources [optional]

Repository: [https://github.com/Ashu-00/NLP-Implementations/tree/main/Decision_Transformer]
Demo: Soon

Uses

Fun, little experiment.

Bias, Risks, and Limitations

This model is biased based on its training data. I take no responsibility for its generation.

Most generated text is non-coherent due to lack of variety of training data.

How to Get Started with the Model


import torch.nn.functional as F

def generate_conditioned_text2(model, tokenizer, prompt, target_rtg, max_length=50, temperature=1.0, top_k=50):
    inputs = tokenizer(prompt, return_tensors="pt")
    input_ids = inputs["input_ids"].to(device)
    attention_mask = inputs["attention_mask"].to(device)

    # Create RTG tensor with the target value for each token in the prompt
    rtg = torch.tensor([[target_rtg] * input_ids.shape[1]], dtype=torch.float).to(device)

    seq_length = input_ids.shape[1]
    for _ in range(max_length):
        with torch.no_grad():
            # Slice rtg to match current sequence length
            rtg_current = rtg[:, :seq_length]
            outputs = model(
                input_ids=input_ids,
                attention_mask=attention_mask,
                rtg=rtg_current,
                return_dict=True
            )

        # Get next token logits and apply temperature scaling
        next_token_logits = outputs["logits"][:, -1, :] / temperature

        # Apply top-k filtering
        top_k_logits, top_k_indices = torch.topk(next_token_logits, top_k)
        probabilities = F.softmax(top_k_logits, dim=-1)
        next_token = top_k_indices[0, torch.multinomial(probabilities, num_samples=1)]

        # Append the predicted token to input_ids and update attention mask

        input_ids = torch.cat([input_ids, next_token], dim=-1)
        attention_mask = torch.cat([attention_mask, torch.ones_like(next_token)], dim=-1)

        # Append the target reward for the new token
        new_rtg = torch.tensor([[target_rtg]], dtype=torch.float).to(device)
        rtg = torch.cat([rtg, new_rtg], dim=1)

        # Stop if EOS token is generated
        if next_token.item() == tokenizer.eos_token_id:
            break

        seq_length += 1

    return tokenizer.decode(input_ids[0], skip_special_tokens=True)

less_toxic_text = generate_conditioned_text2(model, tokenizer, prompt, target_rtg=1)
more_toxic_text = generate_conditioned_text2(model, tokenizer, prompt, target_rtg=0.0)
avg_toxic = generate_conditioned_text2(model,tokenizer, prompt, target_rtg=0.5 )

print("More Toxic Text:", less_toxic_text)
print("Less Toxic Text:", more_toxic_text)
print("Avg Toxic Text:", avg_toxic)

Training Details

Refer to the github for training datasets and procedure.