RuijiaZ/geometry-preserving-orthonormal-init-rlvr-artifacts

Geometry-Preserving Orthonormal Initialization RLVR Artifacts

This repository hosts LoRA adapter checkpoints for the DAPO 1.5B cosine learning-rate runs.

Structure

• —checkpoints/LoRA/seed{41,42,43}/global_step_{50..500}
• —checkpoints/RLMO/seed{41,42,43}/global_step_{50..500}
• —checkpoints/RLPO/seed{41,42,43}/global_step_{50..500}

Each checkpoint directory contains PEFT LoRA adapter files:

• —adapter_config.json
• —adapter_model.safetensors

Base model: deepseek-ai/DeepSeek-R1-Distill-Qwen-1.5B. Code and reproduction scripts: https://github.com/Richard-ZZZ/geometry-preserving-orthonormal-init-rlvr