I just released ZY-S1 Preview, a quick experimental model from ZYR AI, built in India: https://huggingface.co/zyr-AGENT/ZY-S1
Training
- 1.183B parameters, fine-tuned from Sapient AI's HRM-Text-1B (a PrefixLM)
- 1,000 SFT steps plus 120 corrective steps on a filtered MnemicAI/Ling-Coder-SFT-English-Clean
- Apache-2.0
Early scores (25 questions each, base → ZY-S1)
- DROP: EM 4 → 4, F1 18.24 → 19.76
- ARC-Challenge: 76 → 80
- MMLU: 48 → 52
- MATH-500: 4 → 0
Running it
Load it with trust_remote_code=True and pass token_type_ids (1 for prompt tokens, 0 for generated tokens).
Full evals are coming next. Feedback is very welcome! 🇮🇳
Top comments (0)