README.md

metadata

license: cc-by-4.0
language:
  - en
  - fr
pretty_name: AfriTemp-Bench
task_categories:
  - question-answering

AfriTemp-Bench: A Benchmark for Temporal Reasoning over African Statistical and Documentary Data

AfriTemp-Bench is a supervised fine-tuning (SFT) and evaluation dataset for temporal reasoning over African economic indicators. It comprises 12,568 examples across 25 generator families, 9 economic domains, 51/54 AU member states, and 2000--2024.

All labels are deterministic — computed by verifiable arithmetic and comparison functions, not by language models. Approximately 19.6% of examples involve counterfactual or generated worlds; the remainder operate on observed World Bank data. 84.8% include excerpted evidence from World Bank economic reports retrieved via paragraph-level domain-aware scoring.

---

1. Motivation

Temporal reasoning — understanding how measurable quantities change, comparing values across time, and reasoning about trend direction — is a fundamental capability for language models applied to economics, finance, public policy, and data analysis. Existing benchmarks evaluate temporal understanding through synthetic narrative datasets (TempReason, TimeDial, TempLAMA) or news-based QA, but none address the structured, indicator-driven setting of development economics.

AfriTemp-Bench fills this gap by providing:

• Real numerical data from the World Bank's WDI (World Development Indicators) and WGI (Worldwide Governance Indicators), ingested as versioned, immutable snapshots.
• Deterministic gold labels computed by verified arithmetic, not teacher-model distillation.
• Document-grounded context selected from 300+ World Bank economic reports via paragraph-level scoring.
• Curated real-world events (elections, rate decisions, IMF programs, health emergencies) wired into causal and policy-status generators.
• Bitemporal structure: knowledge cutoffs, vintages, and revision edges for evaluating temporal grounding.

---

2. Data Sources

2.1 Indicator Data (WDI / WGI)

All numerical observations are sourced from the World Bank API (wbgapi):

• WDI (database 2): World Development Indicators — annual time series for 54 AU member economies.
• WGI (database 3): Worldwide Governance Indicators — six composite governance scores on a -2.5 to +2.5 scale.

Publication-time honesty: The WB API serves only the current latest value for each indicator-year-entity triple. It provides no per-observation publication timestamps. ATIC records each observation with publication_basis: ingestion_snapshot, meaning the recorded publication time equals the ATIC snapshot time. Real bitemporality (vintage-aware querying) accrues through scheduled immutable snapshot diffs.

2.2 Document Data (WB Documents & Reports API)

Economic reports are harvested from the WDS v2 API (searchdocuments endpoint) using a broad query per country: qterm={country}+economic&rows=50. The harvest pipeline:

1. Searches each of 54 AU countries.
2. Downloads matching PDFs via the WDS download URL template.
3. Extracts text with pdfplumber, normalizing Unicode and stripping watermarks ('dezirohtuA', 'erusolcsiD', 'cilbuP').
4. Filters by _doc_matches_country() using COUNTRY_ALIASES with accent-insensitive matching.
5. Removes garbled documents (empty/short/broken-line threshold, low alpha-density).
6. Scores each document for economic content via keyword density.

The final document corpus spans 366 clean PDFs across all 54 AU countries (573 MB extracted text).

Paragraph-level retrieval: When grounding an example, relevant_documents() selects up to 3 documents closest in publication year to the target year, then scores each paragraph by:

• Domain keyword matches (macro_prices, health, climate, etc.)
• Generic economic signal words
• Numerical density
• Year proximity (decay = 0.02 per year offset)
• Length bonus (80--300 words optimal)

Top paragraphs are concatenated to a 3,000-character budget and prepended to the user prompt as Document excerpts:\n\n[From: {title}]\n{text}.

2.3 Curated Events

98 real-world events across all 54 AU countries were manually curated:

• Elections: National elections with known dates.
• Policy rate decisions: Central bank rate changes (e.g., Bank of Ghana, Central Bank of Nigeria, South African Reserve Bank).
• IMF programs: Arrangements (ECF, EFF, RFI, PLL) with announcement dates and effective intervals.
• Health emergencies: COVID-19 PHEIC and WHO monkeypox PHEIC.
• Regional milestones: AfCFTA trading commencement.

Events are wired into PolicyStatusGenerator (ask whether a policy was active at a query time) and CausalReasoningGenerator (ask whether an event plausibly caused an observed indicator movement).

---

3. Indicators

55 indicator series across 9 domains. Each indicator carries metadata: series_id, name, short_name, domain, unit, measure_type (level/rate/share/index/score), and aggregation (additive/non_additive/average).

Domain	Indicators	Coverage
macro_prices	2666	21.2%
health	1665	13.2%
governance	1280	10.2%
agriculture	1242	9.9%
demography	1199	9.5%
education_labour	1151	9.2%
trade	1077	8.6%
energy	681	5.4%
climate	503	4.0%

series_id	Name	Domain
NY.GDP.MKTP.KD.ZG	GDP growth (annual %)	macro_prices
NY.GDP.PCAP.CD	GDP per capita (current US$)	macro_prices
NY.GDP.MKTP.CD	GDP (current US$)	macro_prices
NY.GDP.MKTP.KD	GDP (constant 2015 US$)	macro_prices
FP.CPI.TOTL.ZG	Inflation, consumer prices (annual %)	macro_prices
PA.NUS.FCRF	Official exchange rate (LCU per US$)	macro_prices
NY.GNP.PCAP.CD	GNI per capita, Atlas method (current US$)	macro_prices
BN.CAB.XOKA.GD.ZS	Current account balance (% of GDP)	macro_prices
NY.GDS.TOTL.ZS	Gross domestic savings (% of GDP)	macro_prices
GC.XPN.TOTL.GD.ZS	Government expenditure (% of GDP)	macro_prices
SP.DYN.LE00.IN	Life expectancy at birth (years)	health
SP.DYN.IMRT.IN	Infant mortality rate (per 1,000)	health
SH.XPD.CHEX.GD.ZS	Health expenditure (% of GDP)	health
SH.IMM.MEAS	Measles immunization (% of children)	health
SH.DYN.MORT	Under-5 mortality rate (per 1,000)	health
SH.STA.MMRT	Maternal mortality ratio (per 100,000)	health
SH.STA.BRTC.ZS	Skilled birth attendance (% of total)	health
SH.HIV.INCD.TL	HIV incidence (per 1,000 uninfected)	health
AG.LND.FRST.ZS	Forest area (% of land area)	climate
EG.FEC.RNEW.ZS	Renewable energy (% of total final)	climate
EG.USE.COMM.FO.ZS	Fossil fuel energy (% of total)	climate
ER.PTD.TOTL.ZS	Protected areas (% of total)	climate
NV.AGR.TOTL.ZS	Agriculture (% of GDP)	agriculture
AG.YLD.CREL.KG	Cereal yield (kg/ha)	agriculture
AG.PRD.CROP.XD	Crop production index (2014-2016=100)	agriculture
AG.LND.ARBL.ZS	Arable land (% of land area)	agriculture
SP.POP.TOTL	Population (total)	demography
SP.POP.GROW	Population growth (annual %)	demography
SP.URB.TOTL.IN.ZS	Urban population (% of total)	demography
SP.POP.65UP.TO.ZS	Population ages 65+ (% of total)	demography
SP.DYN.CBRT.IN	Crude birth rate (per 1,000)	demography
SP.DYN.CDRT.IN	Crude death rate (per 1,000)	demography
GOV_WGI_GE.EST	Government Effectiveness (score)	governance
GOV_WGI_CC.EST	Control of Corruption (score)	governance
GOV_WGI_RL.EST	Rule of Law (score)	governance
GOV_WGI_PV.EST	Political Stability (score)	governance
GOV_WGI_RQ.EST	Regulatory Quality (score)	governance
GOV_WGI_VA.EST	Voice and Accountability (score)	governance
NE.EXP.GNFS.ZS	Exports (% of GDP)	trade
NV.IND.TOTL.ZS	Industry (% of GDP)	trade
NE.TRD.GNFS.ZS	Trade (% of GDP)	trade
BX.KLT.DINV.WD.GD.ZS	FDI net inflows (% of GDP)	trade
NE.IMP.GNFS.ZS	Imports (% of GDP)	trade
TX.VAL.TECH.MF.ZS	High-tech exports (% of manufactured)	trade
EG.ELC.ACCS.ZS	Electricity access (% of population)	energy
IT.NET.USER.ZS	Internet users (% of population)	energy
EG.USE.PCAP.KG.OE	Energy use (kg oil eq. per capita)	energy
EG.IMP.CONS.ZS	Energy imports, net (% of use)	energy
SE.PRM.ENRR	Primary enrollment (% gross)	education_labour
SE.SEC.ENRR	Secondary enrollment (% gross)	education_labour
SL.UEM.TOTL.ZS	Unemployment (% of labor force)	education_labour
SL.TLF.CACT.ZS	Labor force participation (%)	education_labour
SE.ADT.LITR.ZS	Adult literacy rate (% ages 15+)	education_labour
SE.TER.ENRR	Tertiary enrollment (% gross)	education_labour
SL.EMP.TOTL.SP.ZS	Employment-to-population ratio (%)	education_labour

---

4. Geography

All 54 African Union member states, grouped into 5 AU official regions:

Region	ISO3 Codes
North (6)	DZA, EGY, LBY, MRT, MAR, TUN
West (15)	BEN, BFA, CPV, CIV, GMB, GHA, GIN, GNB, LBR, MLI, NER, NGA, SEN, SLE, TGO
Central (8)	CMR, CAF, TCD, COG, COD, GNQ, GAB, STP
East (15)	BDI, COM, DJI, ERI, ETH, KEN, MDG, MUS, RWA, SYC, SOM, SSD, SDN, TZA, UGA
Southern (10)	AGO, BWA, SWZ, LSO, MWI, MOZ, NAM, ZAF, ZMB, ZWE

The Sahrawi Arab Democratic Republic (ESH) is not a WB economy and is excluded. Regional quotas (REGION_QUOTAS) follow AU official proportions: North 0.15, West 0.25, Central 0.15, East 0.25, Southern 0.20.

Fiscal calendar overrides are defined for KEN (July--June), EGY (July--June), ZAF (April--March), and ETH (July 8--July 7).

---

5. Temporal Framework

5.1 Intervals

All temporal extents are modeled as half-open [start, end) intervals with metadata (precision, label, calendar, certainty). Supported normalization patterns:

• "2019" → calendar year
• "Q3 2019" → calendar quarter
• "H2 2021" → half-year
• "2019-05" → calendar month
• "FY 2022/23" → fiscal year (jurisdiction-aware)
• "FY2022" → fiscal year starting in 2022

5.2 Snapshots and Vintages

The WB API is stateless — it always returns the current latest revision. ATIC builds bitemporality through scheduled immutable snapshot pulls:

1. Each snapshot is an independent Parquet file keyed by (series_id, entity_id, year, snapshot_id).
2. Observation IDs are SHA1 hashes of {series_id}|{entity_id}|{year}|{snapshot_id}.
3. The Warehouse class compares consecutive snapshots to detect revision edges (value changes between vintages).

Seeded revisions: seed_revisions.py creates a perturbed prior snapshot at 30% perturbation rate (magnitude ±3%--15%) so that RealVintageGenerator can produce T4 cutoff examples with real revision edges. The current corpus has 1,309 seeded revision edges across 9 series.

5.3 Knowledge Cutoffs

Every example carries a knowledge_cutoff timestamp. Verifiers ensure:

• No evidence published after the cutoff is cited (verify_cutoff).
• Any evidence explicitly excluded must actually be post-cutoff (verify_excluded_evidence).
• Abstention-consistency: if the cutoff makes answer impossible, the model must abstain (verify_abstention_consistency).

---

6. Generator Families

25 generator families produce the 12,568 examples. Each family implements generate(ctx, budget) -> Iterator[TrainingExample] and targets a specific temporal reasoning capability.

Family	Count	Share	Description
cutoff.synthetic_vintage	1500	11.9%	Minimal-pair T4 reasoning across synthetic revision worlds
real.change	1346	10.7%	Absolute difference computation between two years
real.trend	1047	8.3%	Trend classification (increasing/decreasing/stable/non-directional)
real.state_at_time	893	7.1%	Factual retrieval with document grounding
tool.use	749	6.0%	Multi-step tool composition for compound questions
real.yoy	742	5.9%	Single-period relative (year-over-year) change
counterfactual	669	5.3%	Perturb-then-recompute what-if scenarios
events.causal_reasoning	524	4.2%	Temporal alignment of events with indicator movements
real.cagr	450	3.6%	Compound annual growth rate computation
real.rank	450	3.6%	Cross-sectional ordinal ranking among peer countries
spatial.extrema	450	3.6%	Find the year with the highest/lowest indicator value in a window
spatial.first_last	443	3.5%	Identify the first year an indicator exceeded a threshold
spatial.before_after	435	3.5%	Compare average indicator values before vs after a pivot year or event
negative.unit_trap	387	3.1%	Unit/scale mismatch detection
events.policy_status	377	3.0%	Binary classification of policy/event active status at a query date
negative.missingness	324	2.6%	Abstention when observations are absent
cutoff.real_vintage	300	2.4%	Cutoff-aware retrieval over seeded revision edges
periods.normalize	295	2.3%	Fiscal-calendar-aware period expression conversion
dpo.causal	210	1.7%	DPO preference pair for causally-grounded vs. spurious claims
spatial.multi_entity	209	1.7%	Cross-country comparison of the same indicator in the same year
dpo.comparative	196	1.6%	DPO preference pair for correct vs. incorrect comparisons
dpo.cutoff	190	1.5%	DPO preference pair for cutoff-aware vs. cutoff-oblivious answers
dpo.schema	160	1.3%	DPO preference pair for schema-compliant vs. malformed answers
events.causal_evidence	142	1.1%	Document-grounded causal evidence from paragraph co-occurrence
dpo.unit	80	0.6%	DPO preference pair for unit-correct vs. unit-confused answers

6.1 Real-Data Tasks

real.state_at_time: "What was the value of indicator X for country Y in year Z?" — factual retrieval with document grounding.

real.change: "By how much did indicator X change for country Y between year A and year B?" — absolute difference computation.

real.trend: "Was the trend of indicator X for country Y over period P increasing, decreasing, stable, or not directional?" — requires 4+ data points, 70% monotonic share threshold, ±5% relative epsilon.

real.cagr: "What was the CAGR of indicator X for country Y from year A to year B?" — compound annual growth rate calculation.

real.yoy: "What was the year-over-year change of indicator X for country Y in year Z?" — single-period relative change.

real.rank: "Among [countries], where did country Y rank on indicator X in year Z?" — cross-sectional ordinal comparison.

6.2 Temporal Cutoff Tasks

cutoff.synthetic_vintage: Creates two parallel worlds differing only in the knowledge cutoff date. Earlier cutoff sees older observation; later cutoff sees revised value. The gold answers differ between worlds, creating a minimal pair for T4 (cutoff-aware) reasoning.

cutoff.real_vintage: Uses the 1,309 seeded revision edges. Asks "According to a report published at time T1, what was the value of X in year Y for country Z?" — requires reading the correct vintage layer.

6.3 Event-Driven Tasks

events.policy_status: "As of [query date], was [specific policy/event] active?" — requires knowing effective intervals of real-world events.

events.causal_reasoning: "Did [event E] plausibly cause the observed change in indicator X for country Y between T1 and T2?" — temporal alignment of event intervals with indicator movements, with hedging-aware verifiers.

6.4 Negative / Robustness Tasks

negative.missingness: Indicator has no observation for the requested year — model must abstain rather than hallucinate.

negative.unit_trap: Questions with mismatched units (e.g., asking for GDP per capita when given total GDP) — model must detect the mismatch.

6.5 Period Normalization

periods.normalize: Convert between period expressions (e.g., "what was the value in Q3 2019?") using the fiscal calendar and quarter resolution system.

6.6 Counterfactual

counterfactual: "If indicator X had been [value] instead of [actual] in year Y, what would the rank/change/GDP be?" — perturb-then-recompute.

6.7 Tool-Use

tool.use: Multi-step reasoning where the model must select and compose tools (lookup, compute, compare) to answer a compound question.

6.8 DPO Preference Pairs

dpo.cutoff, dpo.schema, dpo.comparative, dpo.causal, dpo.unit: Preference pairs for DPO alignment. Each pair contrasts a correctly-grounded answer against a plausible-but-wrong answer sharing surface form features.

---

7. Verification Pipeline

Every example passes deterministic verification before being written to the output. Verification is not a model-based judge; all checks are computable functions.

Check	Description
verify_calculation	Recomputed every arithmetic operation (identity, absolute change, percent change, CAGR) from input observations. Tolerance: rtol=1e-6, atol=1e-9.
verify_cutoff	No evidence citation has a publication timestamp after the example's knowledge_cutoff.
verify_abstention_consistency	If abstention is set, no calculations or evidenced claims may be present; conversely, if the answer claims a value, abstention must be false.
verify_excluded_evidence	Every item in excluded_evidence must have a publication timestamp strictly after knowledge_cutoff.
trend_classification	For T3_TREND tasks: re-computes classify_trend(observations) and compares to the gold trend_label.

Examples that fail verification are counted (verify_failed) but still included; they are flagged via verification.verified: false and can be filtered downstream.

---

8. Dataset Splits

Splits are group-based, not row-level random. Each example carries a set of group_keys derived from its generator family and entity/synthetic-world identity. The split assignment is deterministic via SHA1 hashing:

bucket = int(hashlib.sha1("|".join(sorted(group_keys)).encode()).hexdigest(), 16) % 1000 / 1000.0
if bucket < 0.08:    benchmark
elif bucket < 0.16:  validation
else:                train

This guarantees that all examples from the same underlying world (e.g., both views of a synthetic-vintage minimal pair) land in the same split, preventing information leakage.

Split	Examples	Share
train	10,727	84.6%
validation	998	7.9%
benchmark	956	7.5%

---

9. Dataset Schema

Each example is a flat dictionary with the following keys:

Field	Type	Description
example_id	string	Unique hash-based identifier
Field	Type	Description
-------	------	-------------
example_id	string	Unique hash-based identifier
system_prompt	string	ATIC system prompt defining the model persona
user_prompt	string	User query, optionally prefixed with document excerpts
gold_answer	string	Deterministically computed answer
generator_family	string	Generator family name (25 total)
task_id	string	Primary task taxonomy (e.g., T2_STATE_AT_TIME, T3_TREND, T5_CAUSAL_EVIDENCE)
synthetic	bool	True if the example is set in a generated/counterfactual world
split	string	train / validation / benchmark
difficulty	int	0--4 ordinal difficulty rating
answer_value	float or null	Numeric answer value (extreme_value, cagr, delta, or null)
answer_year	int or null	Target year of the question
answer_entity	string	Human-readable entity name (e.g., 'Ghana')
answer_entity_id	string	Internal entity ID (e.g., 'geo_gha')
answer_indicator	string	WB series ID (e.g., 'NY.GDP.MKTP.KD.ZG')
answer_unit	string	Unit of measurement
answer_region	string	AU region (west/east/central/southern/north)
answer_country_iso3	string	ISO3 country code
requires_calculation	bool	Whether arithmetic is needed
requires_retrieval	bool	Whether evidence retrieval is needed
requires_reasoning	bool	Whether multi-hop reasoning is needed
num_evidence_items	int	Number of observations in context
num_documents	int	Number of document excerpts provided
num_events	int	Number of curated events in context
sources	list	Provenance source IDs for traceability
observation_ids	list	Source observation IDs

The full internal schema (Pydantic models) includes Observation, Interval, TrainingExample (the universal envelope), Verification, Target, TemporalScope, EventRecord, and DocumentRecord — all defined in atic/models.py.

---

10. Quality Assurance

• Watermark stripping: Removed WB document watermarks ('dezirohtuA', 'erusolcsiD', 'cilbuP') from 252/313 documents. Post-strip watermark artifacts in excerpts: 0%.
• Garbage document filter: Rejects documents with <500 chars, alpha-density <0.35, short-line ratio >0.5, or no economic keyword hits. 22 documents removed.
• Country-match filter: Accent-insensitive alias matching with COUNTRY_ALIASES eliminates false positives (e.g., Turkey/ZAF, Iran/LBY).
• Verification: All 5 deterministric checks pass on every example. Failed examples are flagged, not silently dropped.
• Test suite: 38 pytest tests covering interval arithmetic, trend classification, all generator families, and end-to-end pipeline integrity.

---

11. Limitations

1. Publication-time approximation: WB API statelessness means real vintage discovery depends on snapshot diffs. Some revision edges may be missed between infrequent snapshots.
2. Document quality variance: some countries have only procurement plans or sector reports; quality filter may still reject thin economic content for a few countries.
3. Temporal coverage: Data spans 2000--2024; earlier periods are not represented.
4. Language bias: Documents are primarily English and French; Portuguese-language documents (ANG, MOZ) are fewer.
5. Event depth: 98 events cover all 54 countries, but countries average only 1--2 events each. Event-driven examples may lack variety per country.

---

12. License

12.1 Attribution

Base indicator data sourced from the World Bank's World Development Indicators (WDI) and Worldwide Governance Indicators (WGI). Document corpus sourced from the World Bank Documents & Reports repository.

The World Bank bears no responsibility for the derived labels, task definitions, temporal reasoning framework, or any analysis presented in this dataset.

12.2 Licenses

• WDI data: CC-BY-4.0
• WGI data: CC-BY-NC-3.0-IGO (requires manual review per source terms)
• WB Documents: CC-BY-NC-3.0-IGO
• Curated events and generated labels: CC-BY-4.0

---

13. Citation

@misc{afritemp-bench-2026,
  author = {{Electric Sheep Africa}},
  title = {AfriTemp-Bench: A Benchmark for Temporal Reasoning over African Statistical and Documentary Data},
  year = {2026},
  howpublished = {\url{https://huggingface.co/datasets/africatic/afritemp-bench}},
}