nrl-ai/vn-rag-bench

nom-vn benchmarks

Reproducible benchmark fixtures and baseline results for the nom-vn Vietnamese AI toolkit. Covers retrieval / RAG and Vietnamese OCR; more components arriving in follow-up commits.

Layout

fixtures/                  — input corpora (JSON) and tiny image sets
baselines/                 — bench result JSONs (small, version-controlled)
fixture_builder.py         — Zalo Legal QA sampler (regenerates fixtures)

Why this exists

Per CLAUDE.md principle 12: every metric we publish must come from a committed-and-runnable script with a baseline JSON we can re-measure on every change. This repo holds those baselines and the fixtures that produce them, so anyone can re-run, audit, and compare.

---

Component 1 — Vietnamese RAG retrieval

Source dataset: sampled from GreenNode/zalo-ai-legal-text-retrieval-vn (MIT), itself a HuggingFace mirror of the Zalo AI Challenge 2021 Legal Text Retrieval public corpus (60,701 articles + 788 queries with qrels).

Fixtures

File	Articles	Questions
fixtures/vn_legal_tiny.json	12	12
fixtures/vn_legal_zalo_2k.json	~1.5k	50
fixtures/vn_legal_zalo_5k.json	~5k	80
fixtures/vn_legal_zalo_full.json	~61k	788

(The _full fixture is large; regenerate via fixture_builder.py if the JSON isn't committed in your clone.)

Reproducing the bench

git clone https://github.com/nrl-ai/nom-vn
cd nom-vn
pip install -e ".[chat,otel]" datasets

# Rebuild a fixture:
python benchmarks/rag/fixtures/_build_zalo_legal.py \
    --n-questions 80 --n-distractors 5000 --seed 42 \
    --out benchmarks/rag/fixtures/vn_legal_zalo_5k.json

# Real-models bench, GPU auto-pick:
python benchmarks/rag/bench_rag_vn.py \
    --fixture benchmarks/rag/fixtures/vn_legal_zalo_5k.json \
    --embedder vietnamese \
    --reranker BAAI/bge-reranker-v2-m3 \
    --retrievers bm25,dense,hybrid,hybrid+rerank \
    --device auto \
    --json benchmarks/rag/baselines/zalo_5k__dangvantuan__bge_v2_m3.json

# Whole grid:
bash benchmarks/rag/run_grid.sh

See per-baseline JSON for exact config + metrics.

---

Component 2 — Vietnamese OCR

Source dataset: vn_ocr_subset — 478 image samples deterministically drawn (seed=42) from ducto489/ocr_datasets shard 0 (Apache-2.0), filtered to rows containing Vietnamese diacritics and at least 8 characters of ground-truth text.

The synthetic pencil-rendered fixture (benchmarks/data/synthetic_ocr_vi/, 20 + 20 images) stays in the main repo as a CI smoke test — Tesseract gets ~100% on it which makes it useless for ranking engines.

Reproducing the bench

# Rebuild the VN-only OCR subset (downloads one parquet shard ~150 MB):
python benchmarks/data/vn_ocr_subset/_build.py --n 500 --seed 42

# Run engines:
python benchmarks/accuracy/bench_ocr_real.py \
    --corpus benchmarks/data/vn_ocr_subset \
    --variant none \
    --engines tesseract,easyocr \
    --device cpu \
    --json benchmarks/results/ocr_vn_subset__tesseract_easyocr.json

Engines benchmarked so far:

• Tesseract 5 (vie traineddata, system-installed; Apache-2.0).
• EasyOCR 1.7+ (Apache-2.0; pip-installable).

Engines in the survey but not yet benched:

• VietOCR (Apache-2.0; install broken on Python 3.13 — pinned for follow-up).
• PaddleOCR PP-OCRv5 (Apache-2.0; lightweight detection + recognition).
• Qwen2-VL-2B (Apache-2.0; ~4 GB, GPU-recommended; deferred).
• Surya OCR — GPL-3.0 code + open-RAIL-M models, license-incompatible with our Apache-2.0 default surface, can only bench for comparison.

---

Licenses

This dataset repo: MIT.

Per-corpus licenses:

• fixtures/vn_legal_tiny.json — hand-curated, MIT.
• fixtures/vn_legal_zalo_*.json — derived from GreenNode/zalo-ai-legal-text-retrieval-vn (MIT), itself the Zalo AI Challenge 2021 public corpus.
• vn_ocr_subset/ (when added) — derived from ducto489/ocr_datasets (Apache-2.0).

Bench result JSONs reference embedder / reranker model IDs; consult each upstream model card for license. We standardise on Apache-2.0

• safetensors per CLAUDE.md file-format trust ladder.