olaverse/african-typed-decisions

African Typed Decisions

A benchmark for System One decision models in eight African languages. Each case gives a model one piece of unstructured state (a message, ticket, record or transcript) and asks several typed questions about it. Every answer is a probability distribution over the options the caller supplies:

Type	The model receives	It returns
choice	instructions + named options (criteria: key → description)	a distribution over the keys
noul	a yes/no statement	p(yes)
score	instructions + ordered level descriptions	a distribution over the levels

The schema follows the System One primitives used by LocalLLaMA/typed-decisions, so a row replays against any API or model with that request shape.

from datasets import load_dataset
import json
ds = load_dataset("olaverse/african-typed-decisions", "yor", split="test")       # or "all"
case = ds[0]
state, questions, gold = case["state"], json.loads(case["questions"]), json.loads(case["gold"])

Composition

3,739 cases and 16,881 decisions (5,167 choice, 5,174 noul, 6,540 score).

	native cases	translated cases	native decisions	translated decisions
Amharic (amh)	71	366	172	1830
Hausa (hau)	97	400	236	2000
Igbo (ibo)	100	396	242	1980
Nigerian Pidgin (pcm)	94	292	252	1460
Somali (som)	93	399	241	1995
Swahili (swh)	107	400	280	2000
Yorùbá (yor)	74	380	173	1900
isiZulu (zul)	93	377	235	1885

• translated: the 400 test cases of LocalLLaMA/typed-decisions (customer service, invoice processing, security incidents, agent-trace observability), translated into each language. Free-text fields of the state, the instructions and the option descriptions are translated; field names, IDs, amounts, dates and option keys are left exactly as written, so the original gold answers still apply. The same case appears in every language, so scores compare directly across languages and against the English original.
• native: cases written directly in each language and set in local contexts (banking and BVN, mobile money, ID registration, health centres, markets, farming, transport, telecoms, schools, moderation, scams and more), delivered as WhatsApp messages, SMS, emails, call transcripts, form records and voice-note transcripts.

Columns

Column	Meaning
id	case id (translated ids match the English benchmark's ids)
lang, language	ISO 639-3 code and name
source	translated or native
workflow	workflow (translated) or domain (native)
state	the text or JSON the decision is about
questions	JSON: {question_id: {type, instructions, criteria}}
gold	JSON: the correct answer per question (option key, true/false, or level index)
target	JSON: the reference distribution per question
n_questions	questions in the case

How it was built

1. Translation. A large language model (DeepSeek-V4.1-Flash for Yorùbá, Igbo, Hausa and Pidgin; Qwen3.8-Flash for the others) translated the free-text segments as a list, with instructions to keep every number, amount, date, code, ID and name. A segment set was kept only if it came back complete, every number survived, the length stayed in a plausible range, and lid-lite-608 identified the result as the target language.
2. Native cases. A language model wrote each case with its questions and intended answers (DeepSeek-V4-Pro for Yorùbá, Igbo, Hausa and Pidgin; Qwen3.8-Flash for the others). A second, different model (Qwen3.8-Flash, or DeepSeek-V4.1-Flash for cases Qwen wrote) answered every question independently from the state alone; only questions where the two agreed were kept. target for native questions averages the intended answer with the second model's probabilities.
3. Orthography. Yorùbá is written with full tone marks and underdots, Igbo with dotted vowels; Hausa text had its hooked letters (ɓ ɗ ƙ ƴ) completed by diactag-2.0.

Scoring

Report, per language and per question type: accuracy (top option = gold), Brier score and KL from target, plus ECE for calibration. Send the whole case (state + all questions) in one request, as the English benchmark does, and report the English original alongside for the translated cases.

Leaderboard

Every model below was run through the same harness. Each request contains a whole case: the state and all of its questions. Every answer is turned into a probability vector over the case's options, and every model is scored by the same code. LLMs were prompted for JSON probabilities at temperature 0 with reasoning off. Any answer that is missing or cannot be parsed counts as a uniform guess. The translated column is the main ranking.

#	Model	Size	African, translated ↑	English original ↑	Drop ↓	African, native ↑	KL ↓	Brier ↓	ECE ↓
1	DeepSeek-V4-Pro (prompted)	1.6T	0.669	0.720	0.051	0.955 †	1.125	0.199	0.121
2	PurpleMIST-Flash-1.0	7.9B	0.668	0.720	0.053	0.932 ‡	0.214	0.115	0.049
3	Qwen3.8-Flash (prompted)	undisclosed	0.643	0.712	0.070	0.916 †	0.862	0.216	0.132
4	OpenDecider-small	4B	0.477	0.648	0.171	0.635	0.365	0.199	0.027
5	PurpleMIST-Mini-1.0	1.9B	0.416	0.601	0.185	0.769 ‡	0.399	0.221	0.050
6	laya-multilingual	0.32B	0.322	0.348	0.026	0.280	5.053	0.545	0.360
–	Uniform (same probability on every option)	–	–	0.308	–	–	–	–	–

• Columns:
  • African, translated: accuracy on the 3,010 translated cases (15,050 decisions).
  • English original: accuracy on the same 400 cases in LocalLLaMA/typed-decisions (2,000 decisions).
  • Drop: English minus African.
  • African, native: accuracy on the 729 native cases (1,831 decisions).
  • KL, Brier and ECE are on the translated cases.
  • Size: total parameters; "undisclosed" means the provider does not publish one.
• Rows 1 and 2 are a tie on accuracy: 0.0015 is about 23 of 15,050 decisions. Read KL and Brier to compare the quality of the probabilities.
• † DeepSeek-V4-Pro wrote the Nigerian-language native cases, and Qwen3.8-Flash wrote or checked all of them, so both are scored partly against their own answers in the native column. DeepSeek-V4.1-Flash and Qwen3.8-Flash produced the translations. Translation leaves the gold answers unchanged, but models from those families may find the wording more familiar.
• ‡ PurpleMIST (Flash and Mini) is by the authors of this benchmark. It was trained on African decisions from the same pipeline: other native cases, and translations of other states. None of the cases here were in its training data.

Accuracy by language (translated cases):

Model	Amharic	Hausa	Igbo	Pidgin	Somali	Swahili	Yorùbá	isiZulu
DeepSeek-V4-Pro	0.650	0.675	0.673	0.688	0.678	0.682	0.651	0.657
PurpleMIST-Flash-1.0	0.649	0.649	0.654	0.710	0.683	0.690	0.644	0.671
Qwen3.8-Flash	0.622	0.625	0.627	0.681	0.641	0.661	0.629	0.664
OpenDecider-small	0.491	0.421	0.473	0.619	0.443	0.510	0.439	0.456
PurpleMIST-Mini-1.0	0.330	0.414	0.422	0.589	0.371	0.418	0.420	0.406
laya-multilingual	0.334	0.326	0.317	0.351	0.299	0.311	0.352	0.299

The full metrics are in leaderboard/. To add a model, score it on all with full distributions, sending one request per case, and open a discussion with the numbers.

Notes

• Cases are machine-written and machine-checked, not reviewed by native speakers.
• Nigerian Pidgin has fewer translated cases: translations that read as English were removed.
• Native-case questions were kept only when two models agreed, so they measure clear-cut decisions rather than genuinely ambiguous ones.

Licence

Apache-2.0. The translated cases are derived from LocalLLaMA/typed-decisions (Apache-2.0); please credit it alongside this dataset.

Citation

@misc{african-typed-decisions,
  title  = {African Typed Decisions: a System One decision benchmark in eight African languages},
  author = {Olaverse},
  year   = {2026},
  url    = {https://huggingface.co/datasets/olaverse/african-typed-decisions}
}