Trained a small language model for just generating question (huggingface.co)

🤖 AI Summary
The Tara 1.2 Quest model has been introduced as a specialized language model focused on generating structured question lists in JSON format based on user prompts. Serving as a successor to the earlier Tara 1.1 model, Tara 1.2 Quest is confined to producing exploratory questions relevant to specific topics or short requests, showcasing its ability to respond with a neatly formatted output: {"questions":["...","...","..."]}. This small research model, approximately 5 million parameters in size, leverages the LlamaForCausalLM architecture, offering a context length of 512 tokens and a vocabulary of 4,108. The significance of this model lies in its exploration of constrained question generation, which can enhance user interactions in educational and exploratory contexts. Internal benchmarks demonstrate a strong tendency to produce questions in the expected JSON format and an effective topic-to-question mapping for common subjects, though it struggles with nuanced semantics and can produce repetitive outputs. As a tool primarily for research, Tara 1.2 Quest is not suited for high-stakes applications and is recommended for use as an educational resource rather than a production assistant, given its limitations with negation and off-topic responses.
Loading comments...
loading comments...