| Property | Value |
|---|---|
| License | MIT |
| Training Data Size | 11.5GB |
| Framework | PyTorch, Transformers |
| Authors | DeepESP (Alejandro Oñate Latorre & Jorge Ortiz Fuentes) |
What is gpt2-spanish?
GPT2-Spanish is a specialized language generation model trained from scratch on Spanish text data. It follows the architecture of OpenAI's GPT-2 small model but is specifically optimized for Spanish language generation through both its training data and custom tokenization approach.
Implementation Details
The model was trained using Hugging Face libraries on an Nvidia Tesla V100 GPU. It implements a custom byte-level Byte Pair Encoding (BPE) tokenizer with a vocabulary size of 50,257 tokens, specifically designed to capture Spanish language nuances. The model processes input sequences of 1024 consecutive tokens.
- Custom BPE tokenizer trained specifically for Spanish
- 11.5GB training corpus (3.5GB Wikipedia + 8GB literature)
- Special tokens including "<|endoftext|>" and "<|talk|>"
- Trained on Google Colab servers with V100 GPU
Core Capabilities
- Spanish text generation
- Processing of various literary genres (narrative, poetry, essays)
- Context-aware text completion
- Support for special prompt tokens
Frequently Asked Questions
Q: What makes this model unique?
The model's distinctiveness lies in its ground-up training on Spanish text and custom tokenizer, avoiding the limitations of English-based models adapted to Spanish. The diverse training corpus spanning Wikipedia and literature enables rich linguistic understanding.
Q: What are the recommended use cases?
The model is well-suited for Spanish text generation tasks, including creative writing, content generation, and text completion. It's particularly effective for tasks involving various literary styles due to its diverse training corpus.