Huggggooo/ProtoCycle-7B
ProtoCycle-7B by Huggggooo is a 7.6 billion parameter agentic protein design model, fine-tuned from Qwen2.5-7B-Instruct. This model specializes in multi-step, tool-augmented sequence design, utilizing Group Relative Policy Optimization with Tool-Call Reward (GRPO-TCR) for enhanced performance. It is specifically optimized for complex protein engineering tasks, offering a 32768 token context length for detailed biological sequence manipulation.
Loading preview...
ProtoCycle-7B: Agentic Protein Design Model
ProtoCycle-7B, developed by Huggggooo, is a 7.6 billion parameter model specifically designed for agentic protein design. It represents an RL checkpoint, building upon the SFT checkpoint Huggggooo/ProtoCycle-7B-SFT, which itself was fine-tuned from Qwen/Qwen2.5-7B-Instruct. This model excels at multi-step, tool-augmented sequence design, leveraging a specialized training approach.
Key Capabilities
- Agentic Protein Design: Performs complex, multi-step protein sequence generation.
- Tool-Augmented Design: Integrates external tools for enhanced design capabilities.
- GRPO-TCR Training: Utilizes Group Relative Policy Optimization with Tool-Call Reward for advanced policy learning.
- Specialized Reward Manager: Employs a
proteinreward manager for precise evaluation of design outcomes. - Extended Context: Supports a maximum prompt of 8k tokens and responses up to 20k tokens, facilitating detailed design processes.
Good for
- Protein Engineering: Ideal for researchers and developers working on novel protein sequence design.
- Biotechnology Applications: Suitable for tasks requiring iterative and intelligent design of biological molecules.
- Agent-Based Systems: Demonstrates advanced agentic reasoning and planning for scientific tasks.
ProtoCycle-7B was trained on 10,000 RL prompts from the rl/ subset of the Huggggooo/ProtoCycle-Data dataset, using the VeRL and Open-AgentRL frameworks. Its agent protocol includes explicit <think>, <plan>, <tool_call>, and <answer> tags for structured interaction.