| --- |
| license: apache-2.0 |
| base_model: |
| - google/gemma-4-E2B-it |
| language: |
| - en |
| pipeline_tag: any-to-any |
| library_name: transformers |
| tags: |
| - deep-research |
| - gemma4 |
| --- |
| |
|
|
| # gemma-4-E2B-it — Deep Research Single-Agent RL |
|
|
| Paper: [Learning Self-Orchestration in Decentralized Agent Swarms](https://arxiv.org/abs/2610.22322) |
|
|
| Code: [https://github.com/Embodied-Minds-Lab/MPT](https://github.com/Embodied-Minds-Lab/MPT) |
|
|
| For the three-agent counterpart, see [`Xkev/deepresearch_society`](https://huggingface.co/Xkev/deepresearch_society). |
|
|
| ## Training |
|
|
| - **Initialization**: single-agent SFT cold start of [`google/gemma-4-E2B-it`](https://huggingface.co/google/gemma-4-E2B-it) |
| - **Dataset**: OpenResearcher deep-research tasks |
| - **Framework**: [verl](https://github.com/volcengine/verl) |
| - **Algorithm**: monolithic REINFORCE |
| - **Hyperparameters**: `lr=1e-6`, `lr_warmup_ratio=0`, constant schedule |
|
|
| ## Intended use |
|
|
| Research on multi-agent post-training and deep research agents. Not intended for general dialog or production. |
|
|
| ## License |
|
|
| Apache-2.0. |
|
|
