| language: en | |
| license: apache-2.0 | |
| tags: | |
| - verl | |
| - grpo | |
| - math | |
| - reasoning | |
| - rl | |
| - lora | |
| - peft | |
| base_model: google/gemma-2-9b-it | |
| library_name: peft | |
| # thejaminator/grpo-feature-vector-step-1 | |
| This is a LoRA adapter trained using [verl](https://github.com/volcengine/verl) with GRPO (Group Relative Policy Optimization) | |
| on math reasoning tasks. | |
| ## Training Details | |
| - **Base model**: google/gemma-2-9b-it | |
| - **Framework**: verl GRPO | |
| - **Training steps**: 1 | |
| - **Dataset**: Math reasoning problems | |
| - **Batch size**: 8 | |
| - **Learning rate**: 5e-05 | |
| - **LoRA rank**: 64 | |
| - **LoRA alpha**: 128.0 | |
| - **Number of generations**: 16 | |
| Generated from verl LoRA checkpoint: `/workspace/verl_outputs_feature_vector/global_step_1/actor/lora_adapter` | |