The latest articles on DEV Community by ZYR AI (@zyr_ai). https://dev.to/zyr_ai https://media2.dev.to/dynamic/image/width=90,height=90,fit=cover,gravity=auto,format=auto/https:%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Fuser%2Fprofile_image%2F4181022%2F6501fa01-e12e-4d20-9c99-39a0a8ee306d.png https://dev.to/zyr_ai en ZYR AI Tue, 13 Oct 2026 04:57:32 +0000 https://dev.to/zyr_ai/zy-s1-preview-fine-tuning-a-118b-hrm-text-model-built-in-india-4d6 https://dev.to/zyr_ai/zy-s1-preview-fine-tuning-a-118b-hrm-text-model-built-in-india-4d6 <p>I just released <strong>ZY-S1 Preview</strong>, a quick experimental model from ZYR AI, built in India: <a href="https://huggingface.co/zyr-AGENT/ZY-S1" rel="noopener noreferrer">https://huggingface.co/zyr-AGENT/ZY-S1</a></p> <h2> Training </h2> <ul> <li>1.183B parameters, fine-tuned from Sapient AI's HRM-Text-1B (a PrefixLM)</li> <li>1,000 SFT steps plus 120 corrective steps on a filtered MnemicAI/Ling-Coder-SFT-English-Clean</li> <li>Apache-2.0</li> </ul> <h2> Early scores (25 questions each, base → ZY-S1) </h2> <ul> <li>DROP: EM 4 → 4, F1 18.24 → 19.76</li> <li>ARC-Challenge: 76 → 80</li> <li>MMLU: 48 → 52</li> <li>MATH-500: 4 → 0</li> </ul> <h2> Running it </h2> <p>Load it with <code>trust_remote_code=True</code> and pass <code>token_type_ids</code> (1 for prompt tokens, 0 for generated tokens).</p> <p>Full evals are coming next. Feedback is very welcome! 🇮🇳</p> llm ai machinelearning opensource