This is the GRPO-TCR (Group Relative Policy Optimization with Tool-Call Reward) stage, initialised from the SFT checkpoint Huggggooo/ProtoCycle-7B-SFT.

Base model: Huggggooo/ProtoCycle-7B-SFT (itself fine-tuned from Qwen/Qwen2.5-7B-Instruct)
Training framework: VeRL / Open-AgentRL
Stage: agentic RL with GRPO-TCR
Rollouts per prompt: 8, max turns: 16
Max prompt / response: 8k / 20k tokens
Reward manager: protein (see ProtoCycle/verl/workers/reward_manager/protein.py)

See recipe/protein/reward.py for the exact formulation.

Training Data

10,000 RL prompts for GRPO-TCR training, available at Huggggooo/ProtoCycle-Data (rl/ subset).}

Agent Protocol

<think>  ... reasoning ...  </think>
<plan>   ... stage plan ...  </plan>
<tool_call>{"name": "...", "arguments": {...}}</tool_call>
...
<answer>MAEGEITPLKTF...</answer>

How to Use

See the ProtoCycle repository: ProtoCycle repo.

License

Apache-2.0.

Citation

If you find this work useful, please cite ProtoCycle (forthcoming) and the upstream frameworks: VeRL, Open-AgentRL, ProTrek, ESM.