(cache)anti-language.ipynb

Notebook

Gemini

Run this once to load the model and tokenizer.

Gemini

Type the prompt and run this block. You can play with the max_new_tokens and bottom_k.

Gemini

prompt = "This is "
input_ids = tokenizer.encode(prompt, return_tensors='pt')
max_new_tokens = 50  # how many tokens to generate
with torch.no_grad():
  for _ in range(max_new_tokens):
      # 1. Get the model's predictions for the next token
      outputs = model(input_ids=input_ids)
      next_token_logits = outputs.logits[0, -1, :]  # logits for the last token
      # 2. Convert logits to probabilities
      probs = torch.softmax(next_token_logits, dim=-1)
      # 3. Find the least likely tokens (max. bottom_k pro gpt2 = 50257)
      bottom_k = 5
      bottom_probs, bottom_indices = torch.topk(probs, bottom_k, largest=False)
      # Normalize the probabilities of the bottom tokens so they sum to 1
      bottom_probs = bottom_probs / bottom_probs.sum()
      # 4. Randomly select one of the bottom tokens
      chosen_index = torch.multinomial(bottom_probs, num_samples=1)
      next_token_id = bottom_indices[chosen_index]
      # 5. Append the selected token to the sequence
      input_ids = torch.cat([input_ids, next_token_id.unsqueeze(0)], dim=-1)
# Decode the full sequence
generated_text = tokenizer.decode(input_ids.squeeze(), skip_special_tokens=True)
print(generated_text)

Gemini

Run this block if you want to save the last output as a text file.

Gemini


# Save the raw generated text to a UTF-8 file
with open("output.txt", "w", encoding="utf-8") as f:
    f.write(generated_text)

Variables

Terminal