Multilingual jailbreak attacks on Booster-defended LLMs (Mistral-7B) — French, Yoruba, Yoruba code-switch attack variants with ASR evaluation
# Multilingual Jailbreak Attacks on Booster-Defended LLMs
Tests robustness of **Booster** safety alignment against fine-tuning attacks using multilingual unsafe data — French (high-resource), Yoruba (low-resource), and Yoruba code-switched English. Base model: Mistral-7B-v0.1.
## Key Results
| Variant | ASR | vs. M1 baseline |
|---|---|---|
| Raw (no defense) | 67.4% | — |
| M1 (Booster defended) | 0.9% | baseline |
| M2 French attack | **21.3%** | +23.7× |
| M2 Yoruba code-switch | 3.6% | +4× |
| M2 Yoruba pure | 1.6% | +1.8× |
Booster cuts ASR 67.4% → 0.9% (−98.7%). French attack most effective — high-resource language, strong Mistral-7B representation. Full breakdown: `results/RESULTS_SUMMARY.md`
---
## Reproducing Results
### Prerequisites
- Mistral-7B-v0.1 downloaded to `/workspace/raw_model/` (`mistralai/Mistral-7B-v0.1` on HuggingFace)
- Booster repo cloned to `/workspace/Booster/` (`git clone
github.com`)
- CUDA GPU ≥40GB VRAM (tested on A100)
- Python deps: `pip install -r /workspace/Booster/booster_pip.txt`
### 1. Configure paths
```bash
cp config.example.sh config.sh
# Edit config.sh: set BOOSTER_DIR, OUTPUT_DIR, RAW_MODEL_PATH, HUGGINGFACE_TOKEN
source config.sh
```
### 2. Generate attack datasets (skip if using provided `data/` files)
```bash
# Verify setup first
python3 scripts/test_translation.py # NLLB smoke test
python3 scripts/test_code_switch.py # spaCy span extraction test
# Translate
python3 scripts/translate_french.py # → beavertails_30k_train_french.jsonl
python3 scripts/translate_yoruba.py # → beavertails_30k_train_yoruba.jsonl
python3 scripts/code_switch_yoruba.py # → beavertails_30k_train_codeswitch_yoruba.jsonl
# Convert to instruction format
python3 scripts/convert_data.py --input --output data/beavertails_french_instruct.json
python3 scripts/convert_data.py --input --output data/beavertails_yoruba_instruct.json
python3 scripts/convert_data.py --input - …