kunjcr2/code-adaptroute-v2

code-adaptroute

LoRA adapter for the code domain in AdaptRoute.

Mounted onto a frozen 4-bit NF4 quantised Qwen/Qwen2.5-1.5B at inference time via peft.add_weighted_adapter() — weights provided by the gating network.

LoRA Config

• r = 16, alpha = 32, dropout = 0.05
• Target modules: ['q_proj', 'k_proj', 'v_proj', 'o_proj']
• Training: 2 epochs on 20000 samples, lr=0.0002

Training Data

• Source: iamtarun/python_code_instructions_18k_alpaca