Multilingual AI Guardrail Fine-Tuning

Fine-tuning pipelines for multilingual content-safety guardrails — PII detection, content moderation, prompt-injection detection — across multiple model architectures and languages.

Currently ships Korean + English. Hindi, Tamil, English-Hindi, and more on the roadmap.

Models

All three pipelines share the same dataset format and can be retargeted to any language by pointing DATASET_DIR at a different corpus.

Benchmarks (held-out test split — no train/test leak)

Full per-dataset breakdown in benchmark-results/*.json.

Quick start

# 1. Clone external dataset repos (not committed; see DATA.md)
git clone https://github.com/<org>/korean-guardrail-dataset.git
# etc.

# 2. Point datasets/ at the data you want to train on
mkdir -p datasets
ln -s ../korean-guardrail-dataset/data datasets/korean

# 3. Build training data
cd finetune-gemma
docker compose run --rm format-data
docker compose run --rm prepare-train-data

# 4. Train
docker compose up train

# 5. Serve with vLLM
docker compose up -d vllm

# 6. Benchmark
docker compose up benchmark

Adding a new language

The pipeline is language-agnostic. To add a new language (e.g. Hindi):

1. Prepare a dataset repo with the same schema as korean-guardrail-dataset:

   <your-hindi-repo>/data/
   ├── processed/                # raw per-source JSONL, one per dataset
   └── processed_split/          # train/test splits (run split_datasets.py)
   

   Each record:

   {
     "query": "...",
     "blocked": true,
     "type": "pii-filter | moderation | safety-classifier | rules-based-protections | output-validation",
     "answer": [{"form": "...", "label": "..."}],
     "topic": ["..."]
   }
   

2. Symlink it into datasets/:

   ln -s ../hindi-guardrail-dataset/data datasets/hindi
   

3. Train by pointing DATASET_DIR at it:

   cd finetune-gemma
   DATASET_DIR=../datasets/hindi docker compose run --rm format-data
   DATASET_DIR=../datasets/hindi docker compose up train
   

The same three pipelines (finetune-gemma, finetune-qwen, gliner2_korean_pii) train on any language that follows the schema — no code changes required.

Folder layout

guardrail/
├── README.md                   # this file
├── DATA.md                     # data pipeline / rebuild instructions
├── .gitignore
├── datasets/                   # per-language dataset mountpoints (symlinks, gitignored)
│   └── korean → ../korean-guardrail-dataset/data
├── finetune-gemma/             # Gemma 4 E2B fine-tuning pipeline
├── finetune-qwen/              # Qwen 3.5 0.8B fine-tuning pipeline
├── gliner2_korean_pii/         # GLiNER2 NER pipeline (PII-only)
└── benchmark-results/          # honest held-out F1 scores (committed)

Each finetune-*/ project is self-contained:

finetune-<model>/
├── Dockerfile                  # training environment
├── Dockerfile.vllm             # inference serving environment
├── docker-compose.yml          # format-data / prepare / train / vllm / benchmark services
└── src/
    ├── format_training_data.py # dataset → chat-template JSONL
    ├── prepare_dataset.py      # tokenize, filter, cache
    ├── training_data.py        # shared helpers
    ├── train.py                # TRL + Unsloth fine-tuning
    ├── merge_and_push.py       # merge LoRA, upload to HF
    ├── benchmark.py            # concurrent vLLM API benchmark
    └── inference.py            # ad-hoc testing

Roadmap

• Korean (Gemma 4 E2B, Qwen 3.5 0.8B, GLiNER2)
• English (implicit in both models via multilingual pretraining + English PII data)
• Hindi
• Tamil
• English + Hindi code-switching
• Adversarial augmentation (char obfuscation, homoglyphs, indirect injection — see AccuKnox blog)
• Multimodal (image/PDF prompt injection) for Gemma's vision tower

Key design principles

• Language-agnostic pipelines — same Docker setup, swap DATASET_DIR
• Honest evaluation — deterministic 90/10 train/test split, no leak
• Reproducibility — seeded splits, pinned Docker images, all configs versioned
• Model diversity — three model families covering different compute/accuracy trade-offs

See also

• DATA.md — how to rebuild training data from scratch
• benchmark-results/ — per-dataset F1 scores
• finetune-gemma/docs/ — Gemma-specific plan and notes