Model Card for Infinitode/TWNGM-OPEN-ARC

Repository: https://github.com/Infinitode/OPEN-ARC/

Model Description

OPEN-ARC-TWNG is a simple recurrent neural network (RNN) language model developed as part of Infinitode’s OPEN-ARC initiative. It predicts the next token in a sequence using a lightweight architecture suitable for smaller datasets.

Architecture:

• Embedding: Embedding(vocab_size, 500) maps tokens to 500-dimensional vectors.
• Recurrent Layer: SimpleRNN(50) processes sequences with 50 recurrent units.
• Output Layer: Dense(vocab_size, activation='softmax') produces a probability distribution over the vocabulary.
• Framework: TensorFlow 2.x / Keras
• Training Setup: Compiled with loss='categorical_crossentropy', optimizer='adam', and tracked accuracy metric.

Uses

• Generating creative and fantasy-style weapon or item names.
• Supporting game design workflows and ideation.
• Creating randomized content for prototyping or mods.

Limitations

• May produce implausible or inappropriate names if poorly seeded or prompted.
• Vocabulary is tailored to Terraria-style names; may not generalize to other genres.
• Does not enforce real-world constraints like lore consistency, cultural appropriateness, or game balance.

Training Data

• Dataset: All Terraria Weapons DPS v1.449 dataset from Kaggle.
• Source URL: https://www.kaggle.com/datasets/acr1209/all-terraria-weapons-dps-v-1449
• Content: Weapon names and their damage-per-second values used as creative seed text.
• Size: 395 entries of unique weapon names.
• Preprocessing: Names tokenized into subword units using SentencePiece BPE.

Training Procedure

• Tokenizer: SentencePiece BPE trained on the weapon name corpus (vocab size: 500).
• Batch Size: 128
• Sequence Length: max_seq_len (based on longest tokenized name, 11)
• Optimizer: Adam
• Loss: categorical_crossentropy
• Metrics: accuracy
• Epochs: 500
• Train/Validation Split: 100% train, 0% validation.

Evaluation Results

Metric	Value
Train Accuracy	78.6%
Validation Accuracy	not used
Loss (final)	0.44

How to Use

import random

def generate_random_name(min_length=3, max_length=10, temperature=1.0, seed_text=""):
    # Use a random seed
    random.seed()
    
    if seed_text:
        # If seed text is provided
        generated_name = seed_text
    else:
        # Randomly select a token from our vocab as our starting token if no seed text is present
        random_index = random.randint(1, vocab_size-1)
        random_token = sp.id_to_piece(random_index)
        generated_name = random_token

    # Generate subsequent subword tokens
    for _ in range(max_length - 1):
        # Encode our starting text
        token_list = sp.encode_as_ids(generated_name)
        token_list = pad_sequences([token_list], maxlen=max_seq_len-1, padding='pre')
        
        # Run prediction
        predicted = model.predict(token_list, verbose=0)[0]

        # Apply temperature to predictions, helps to varied results
        predicted = np.log(predicted + 1e-8) / temperature
        predicted = np.exp(predicted) / np.sum(np.exp(predicted))

        # Sample from the distribution
        next_index = np.random.choice(range(vocab_size), p=predicted)
        next_index = int(next_index)
        next_token = sp.id_to_piece(next_index)

        # Add the predicted token to our output
        generated_name += next_token
        
        # Decode the generated subword tokens into a string
        decoded_name = sp.decode_pieces(generated_name.split())

        # Stop if end token is predicted (optional, based on your dataset), or stop if max_length is reached
        if next_token == '' or len(decoded_name) > max_length:
            break

    # Replace underscores with spaces
    decoded_name = decoded_name.replace("▁", " ")
    
    # Remove stop tokens from the output
    decoded_name = decoded_name.replace("</s>", "")
    
    # Capatilize the first letter of each word
    generated_name = decoded_name.rsplit(' ', 1)[0]
    generated_name = generated_name[0].upper() + generated_name[1:]

    # Split the name and check the last part, make sure that it is not cut off
    parts = generated_name.split()
    if parts and len(parts[-1]) < min_length:
        generated_name = " ".join(parts[:-1])
    
    # Strip the output to ensure no extra whitespace
    return generated_name.strip()

Contact

For questions or issues, open a GitHub issue or reach out at https://infinitode.netlify.app/forms/contact.