tw-rag-bench

RAGAS-style retrieval QA over Taiwanese passages — 2,665 items in the RAGAS contract (question / contexts / reference), evaluated with Twinkle Eval's native ragas mode: the system under test answers free-form from the retrieved contexts, and a judge scores faithfulness, answer relevancy, context precision and context recall.

Part of OpenTWBench. Questions derive from tw-formosa-bench; passages from Taiwan.md.

The retrieval setting

Each item hands the system what a mediocre retriever would: the gold passage mixed with three same-category distractors — topically plausible, factually useless. context_precision therefore has a known target (exactly one of four passages is relevant), and the gold's slot (gold_position, drawn from the item's hash; spread 641/648/717/659) makes position bias measurable from the same run.

Running it (Twinkle Eval ragas mode)

# 1. your RAG pipeline answers free-form (or use the reference script)
export OPENAI_BASE_URL=... OPENAI_API_KEY=...
python code/run_rag_generation.py --model <system-under-test> --out responses.jsonl

# 2. assemble the judge set — question field becomes the judge prompt
python code/make_ragas_dataset.py --responses responses.jsonl --out eval/rag.jsonl

# 3. judge with Twinkle Eval (config model = the judge)
twinkle-eval --config ragas.yaml     # evaluation_method: "ragas"

The judge prompt requests the four metrics as strict JSON, which Twinkle's ragas extractor parses unmodified. In the end-to-end smoke the judge returned context_precision: 0.25 — recognising that exactly one of the four passages is relevant, the design working as intended.

Fields (ragas.jsonl, primary)

Field	Description
question	The bare question, Traditional Chinese
contexts	Four retrieved passages, gold among distractors
reference	Declarative reference answer, mechanically composed — 台灣原住民族官方認定的族群數量：16。
gold_position	Which context (1–4) holds the answer
category / source_chunk	Provenance into tw-formosa-bench

Construction guarantees

Zero generation anywhere in the data: questions and passages existed in tw-formosa-bench (whose grounding check verified each answer against its passage), and references are template-composed. Three invariants verified over all 2,665 items: the gold passage sits at its recorded position; no distractor comes from the gold's own article (an adjacent window can paraphrase the fact past a substring check); no distractor contains the answer string (148 candidates rejected).

Secondary config: mcq

The same items as four-option questions with all passages in the prompt — deterministic box scoring, no judge, pairable one-to-one with the closed-book tw-formosa-bench for acc(open) − acc(closed) per question. Use it when you need judge-free numbers; the RAGAS form is the primary evaluation.

Licence

CC BY-SA 4.0, inherited from Taiwan.md through tw-formosa-bench.

Citation

@misc{opentwbench2026,
  title  = {OpenTWBench: An Open Evaluation Suite for Taiwan-Domain Knowledge},
  author = {Huang, Liang-Hsun},
  year   = {2026},
  url    = {https://opentwbench.ai}
}