---
language: ar
license: apache-2.0
base_model: Qwen/Qwen3-8B
tags:
- arabic
- grpo
- activation-steering
- lora
---

# AraSteer: Activation Steering + GRPO for Arabic

Two LoRA adapters trained with Group Relative Policy Optimization (GRPO) on Qwen3-8B
for Arabic language generation improvement.

## Adapters

### grpo_a/
- **Method**: Raw GRPO (200 steps, r=8, 21.8M params)
- **Reward improvement**: +4.8% relative over 200 steps

### grpo_b/
- **Method**: CLAS-warm-started GRPO (200 steps, r=16, 43.6M params)
- **CLAS config**: alpha=1.25, top-4 Arabic-specific layers {34, 33, 32, 0}
- **Reward improvement**: +9.1% relative, +15.9% faster convergence vs GRPO-A at step 50

## Usage



## Paper

AraSteer: Bimodal Neuron Specialization and Activation Steering for Arabic in Multilingual LLMs