| --- |
| license: cc-by-nc-4.0 |
| language: [en, pcm, yo, ha, ig] |
| tags: [safety, elections, nigeria, benchmark, red-teaming, code-mixed, lexsi-hackathon] |
| pretty_name: Ìdìbò election-safety benchmark and data |
| --- |
| |
| # Ìdìbò: election-safety benchmark, training data and results (Nigeria 2027) |
|
|
| Everything behind [`GOPAE/idibo-tiny-aya-earth`](https://huggingface.co/GOPAE/idibo-tiny-aya-earth). |
|
|
| ## Contents |
|
|
| | Path | What | |
| |---|---| |
| | `bench/{en,pcm,yo,ha,ig}.jsonl` | **Benchmark v1**: 109 items per language. 20 fact (with gold answer and source id), 8 abstain (unknowable or conflicting), 18 harmful requests each paired with 18 benign look-alikes, 45 neutrality swaps (6 parties, 5 ethnic groups, 2 religions). Non-English: machine-translated (Cohere models), `src_en` holds the original | |
| | `bench2/en.jsonl` | **Benchmark v2 (held out, written after training)**: 40 harmful/benign pairs over 10 harm categories, 16 jailbreak-wrapped harmful requests (role-play, fiction, research, translation, hypothetical, persona split, JSON, continuation, emotional appeal, …) with 8 benign wrappers, 12 religion swaps | |
| | `bench2/mix.jsonl` | 30 **code-mixed** prompts written by hand (Pidgin-English, Yorùbá-English, Hausa-English), with `src_en` | |
| | `bench2/leakage.json` | Nearest training question for each v2 item (max cosine 0.948; none ≥ 0.95) | |
| | `kb/chunks.jsonl` | 309 cited passages: Electoral Act 2026 sections and INEC voter pages, with URL and date | |
| | `train/{idibo,civic}.jsonl` | Training sets (prompt with golden plus distractor passages → completion) | |
| | `sft/` | English SFT data, the CuratorKIT run (manifest, dataset card, rejected rows) and translations | |
| | `runs/<model>/<lang>.jsonl`, `runs2/…` | Every model answer (v1 and v2), with grades: `.graded.jsonl` (command-a-plus, round 1) and `.graded.local.jsonl` (back-translation by tiny-aya-global, grading by c4ai-command-r7b) | |
| | `results/` | Final summaries with 95% intervals, grader agreement, AuditKIT significance tests, invented-contact counts | |
|
|
| Item fields: `id, kind (fact|abstain|harmful|benign|neutral), category, prompt, expected, gold?, cite?, pair?, swap?, src_en?`. |
|
|
| ## Intended use |
|
|
| Measuring whether an assistant for elections, especially in Nigerian languages: |
| - helps with election harm, or over-refuses look-alike civic questions; |
| - takes political sides; |
| - invents facts. |
|
|
| Harmful items are **requests only**. Model answers to harmful requests are kept only when they refuse: 394 answers that were graded harmful, or that did not open with a refusal, have their text (and its back-translation) replaced with a `[redacted …]` note, with `redacted: true` and the grade kept (`scripts/package_hf_dataset.py`). The dataset contains no harmful completions. |
|
|
| ## Known issues |
|
|
| - Machine translations are unreviewed in places. Example: one Yorùbá fact item asks "where" instead of "when". |
| - We tried to machine-translate v2 with a safety-tuned model (Tiny Aya Earth). It answered or refused many harmful |
| prompts instead of translating them, so v2 is English plus hand-written code-mixed only. |
| - Small: read the 95% intervals the scorer reports. |
| - The open grader (command-r7b) is lenient on harmful items (κ = 0.28 against command-a-plus): harmful-compliance rates it reports are lower bounds. |
|
|
