jtlevine/lastmile-bench-inference

LastMileBench — Public Inference View

The authoritative input for submissions to the LastMileBench evaluator. Each panel from the full private dataset is mirrored here with held-out label fields stripped via a single INFERENCE_WHITELIST source of truth. If you are scoring a model against LastMileBench, download panels from this dataset.

• Site: lastmilebench.jeff-levine.com
• Submit a model: jtlevine/lastmile-bench-submit (Gradio Space)
• Code: github.com/jtlevine18/lastmile-bench

What this benchmark evaluates

AI systems that advise smallholder farmers, commodity traders, and urban informal workers increasingly operate in domains where randomized evaluation is impossible — you cannot randomize a weather system, a price crash, or a heat wave. But these events already happened. LastMileBench replays them and scores AI tools on the decisions they would have made, in units that matter to the end user — Kenyan shillings, millimeters of rain, dollars of insurance payout — not academic accuracy metrics.

How this is structured (and the pattern you may recognize)

LastMileBench follows the same protocol + reference benchmarks + framework pattern as BIG-bench, EleutherAI's lm-evaluation-harness, and OpenAI Evals. A small set of reference benchmarks is maintained by the core team; a framework primitive (DecisionBenchmark) lets anyone stand up a new benchmark in a new domain, run the harness locally, and optionally contribute the panel back via GitHub PR.

• If your model targets one of the reference domains below, submit to the matching panel via the Submit Space.
• If your model targets a domain not on this list (e.g. cassava in Benin, heat exposure in Mombasa, millet in the Sahel), the framework is the path — see "Contribute a new panel" at the bottom of this README.

Panels available here

Path	Benchmark	DPs	Decision	Cost unit
market_intelligence/kenya_maize_daily_v0_1/	market_intelligence	4,506	hold vs. sell over 7 / 14 / 30 days	KES per 100 kg maize
market_intelligence/india_pulses_v0_3/	market_intelligence	16,373	hold vs. sell (pulses)	INR per quintal
market_intelligence/kenya_maize_monthly_v0_2/	market_intelligence	170	monthly hold vs. sell	KES per 90 kg bag
weather_advisory/kerala_tn_advisory_v0_2/	weather_advisory	1,768	protect vs. proceed (per-station calibrated thresholds)	mm rainfall (2-column)
weather_advisory/kerala_tn_advisory_v0_1/	weather_advisory	1,768	protect vs. proceed (legacy fixed 10mm; sensitivity comparison)	mm rainfall (2-column)
climate_risk_insurance/dar_es_salaam_insurance_era5_v0_1/	climate_risk_insurance	132	no trigger / alert / full payout	USD (2-column)

How to submit

from huggingface_hub import hf_hub_download
import json

path = hf_hub_download(
    repo_id="jtlevine/lastmile-bench-inference",
    filename="market_intelligence/kenya_maize_daily_v0_1/decision_points.jsonl",
    repo_type="dataset",
)
decision_points = [json.loads(line) for line in open(path) if line.strip()]

# Run your model and produce one prediction per decision_point_id:
predictions = [
    {"decision_point_id": dp["id"], "recommended_action": "hold_7d", "confidence": 0.8}
    for dp in decision_points
]

# Save to predictions.jsonl and upload at the Submit Space.
with open("predictions.jsonl", "w") as f:
    for p in predictions:
        f.write(json.dumps(p) + "\n")

Each panel ships an INFERENCE_SCHEMA.json listing the exact fields available at inference time. Predictions must cover every decision point in the panel; the scorer raises on coverage mismatch and on SHA-256 mismatch against the panel manifest.

Submissions are scored in seconds and return a Markdown receipt with hit rate and cost metrics in the units of that benchmark.

Why decisions, not forecasts

On the same 104 decision points from the 2015–16 Indian pulse crisis, a popular time-series forecast scored 11.1% MAPE — within the normal range for commodity forecasting — while its decision hit rate was 7.7%. A naive fixed-action baseline beats it 5×. The forecast was close enough in absolute terms to satisfy accuracy metrics but systematically on the wrong side of the decision boundary. That's the empirical case for scoring decisions, not forecasts. See the preprint §5.1 for the mechanism analysis.

Contribute a new panel

The DecisionBenchmark framework accepts any new benchmark instance you define — action space, decide-and-cost function, decision-point schema — and runs it through the same scoring harness that powers the five reference panels above. The examples/fisheries/ directory in the repo is a minimal worked instantiation you can crib from.

Same mechanics as BIG-bench / lm-eval-harness / OpenAI Evals:

data/benchmark/<benchmark>/panels/<your_panel_id>/
├── events.jsonl                 # curated tail events with external citations
├── decision_points.jsonl        # decision points with held-out label fields
├── manifest.json                # SHA-256 hashes, license, attribution, constants
└── README.md                    # dataset card

scripts/build_<benchmark>_<your_panel_id>.py    # reproducible build script

Fork the GitHub repo, add a panel directory + build script, run pytest tests/test_protocol_validation.py to verify protocol compliance, open a PR. Maintainers review against the per-benchmark PROTOCOL.md; merged panels land on the public leaderboard for future submitters. Full quality gates in CONTRIBUTING.md.

License & attribution

This inference view is released under CC-BY-4.0. Upstream data sources have their own licenses; each panel's manifest.json::source_attribution has the required attribution string. Upstream chain:

• market_intelligence (Kenya daily) — KAMIS (Kenya Ministry of Agriculture, Livestock and Fisheries)
• market_intelligence (Kenya monthly) — WFP VAM (CC-BY-IGO)
• market_intelligence (India) — GODL-India via iancovert/Agmarknet GitHub mirror
• weather_advisory — CHIRPS v2 (UCSB / USGS, public domain)
• climate_risk_insurance — ERA5-Land (Copernicus, CC-BY compatible)

Citation

@software{lastmile_bench,
  title   = {LastMileBench: Counterfactual Decision Benchmarks for Last-Mile AI},
  author  = {Levine, Jeff},
  year    = {2026},
  url     = {https://github.com/jtlevine18/lastmile-bench}
}