PatSnap/Hiro-Pharma

Model Overview

Hiro-Pharma is a 7.6 billion parameter causal language model, built upon the Qwen/Qwen2.5-7B base model. It is released in safetensors format under the Apache 2.0 license. This model has been specifically fine-tuned using internally generated synthetic biomedical-domain dialogue data, including objective-question and question-answer examples. The training involved supervised fine-tuning (SFT) followed by reinforcement learning (RL), primarily utilizing Group Relative Policy Optimization (GRPO) with rule-based and model-based rewards.

Key Capabilities

• Biomedical Dialogue: Excels in conversational assistance within the biomedical and pharmaceutical domains.
• Text Generation: Capable of summarization, rewriting, and explanation of biomedical text.
• Research & Development: Intended for evaluation and engineering of biomedical dialogue systems.
• Internal Productivity: Supports exploration of biomedical literature, terminology, and workflows for non-clinical purposes.

Intended Use Cases

• Biomedical and pharmaceutical conversational AI.
• Summarizing and explaining complex biomedical documents.
• Evaluating and developing new biomedical dialogue systems.

Important Considerations

Users should be aware that Hiro-Pharma is not intended for clinical diagnosis, treatment planning, or patient-specific medical advice. It may generate inaccurate or outdated information and is not validated as a medical device. Evaluation on specific biomedical tasks, terminology accuracy, and safety requirements is crucial before production deployment. The model's context length can extend up to 131,072 tokens, depending on runtime and hardware.