R2-Guard Reproduction

This repository contains my reproducibility work for the paper R2-Guard: Robust Reasoning Enabled LLM Guardrail.

The original code release is from:

https://github.com/kangmintong/R-2-Guard

I kept the original project structure and added only the files needed to set up the environment, run the main reproduced experiments, and inspect the results.

What is included

• REPRODUCIBILITY.md: setup instructions, commands used, and notes about what was/was not reproduced.
• repro_table2.sh: local Table 2 style runs.
• repro_weights.sh: weight-learning experiments.
• run_openai_mod_balanced_n64.sh: small OpenAI Moderation subset run used when full API scoring was rate-limited.
• analyze_reasoning_extension.py: reasoning-layer extension analysis using saved result CSVs.
• results/: final CSV/Markdown result summaries kept for the report.

Main result files

results/table2_local_full.csv
results/table2_local_full_deviation.md
results/table2_llamaguard_8bit_full.csv
results/table2_llamaguard_8bit_full_deviation.md
results/table2_official_style_balanced_n64.csv
results/table2_official_style_balanced_n64_deviation.md
results/extension_reasoning_ablation.csv

Quick start

Create a .env file if you want to run gated/API models:

cp .env.template .env

For local Docker runs on my RTX 5070 setup:

docker build -f Dockerfile.local5070 -t r2guard-local5070 .

Then run the local reproduction commands:

bash repro_table2.sh

To reproduce the reasoning-layer extension table from the saved result CSVs:

python analyze_reasoning_extension.py

This creates:

results/extension_reasoning_ablation.csv

The exact full paper reproduction requires external services and stronger hardware than my local machine. Details are in REPRODUCIBILITY.md.