Hello, Hugging Face community!
I’m excited to share ZY-S1 Preview, an experimental language model from the ZYR project, which is being developed in India.
Model: zyr-AGENT/ZY-S1 · Hugging Face ZY-S1 Preview is built on Sapient AI’s HRM-Text-1B base model. It explores additional supervised fine-tuning and corrective fine-tuning to evaluate performance across reasoning, reading comprehension, mathematics, and general knowledge.
In an initial sample of 100 questions across five benchmark metrics, ZY-S1 Preview achieved:
ARC-Challenge: 80% MMLU: 52%
DROP Token F1: 19.76%
DROP Exact Match: 4%
MATH-500 Exact Accuracy: 0% These are preliminary results from small samples, not full benchmark scores. Mathematical reasoning and reading comprehension remain important areas for improvement.
Born in India. Built for Intelligence.
The ZYR project is being developed in India, with a focus on AI model training, experimentation, and evaluation.
ZY-S1 Preview is an early experimental checkpoint—not a finished model or a claim of state-of-the-art performance.
I’d love to hear from the Hugging Face community:
Which benchmarks should I run next?
What improvements would you recommend for future training?
Which reasoning or coding tasks should I prioritize?
Thank you for checking out the project!
ZY-S1 Preview — Born in India. Built for Intelligence.