README.md

---
license: apache-2.0
pretty_name: OpenRelief — Food-Security Forecasting Dataset
task_categories:
- time-series-forecasting
- text-generation
tags:
- food-security
- ipc
- famine-early-warning
- time-series
- humanitarian
- llm-annotation
language:
- en
size_categories:
- 10K<n<100K
---

# OpenRelief

A retrospective research dataset and benchmark pairing district-level food-security
histories with national context, built to study **FEWS NET IPC phase forecasting**
and **time-series-to-language supervision** — teaching a language model to connect
historical signals into a written, evidence-cited rationale rather than just a number.

Built for the **Temporal AI Challenge**. Source code, frontend and full documentation:
[github.com/luk-huebner/OpenRelief](https://github.com/luk-huebner/OpenRelief).

## Task

Given six months of history across 21 channels (food-security indices, prior IPC
assessments, shipping, conflict, rainfall, staple prices), predict the FEWS NET IPC
phase (1 Minimal → 5 Famine) three calendar months ahead, with a generated rationale,
cross-domain hypotheses and recommended actions.

## Dataset at a glance

| Field | Value |
|---|---|
| Dataset version | `8142c89de89cb862117d6a814be515b9f00312bb960e33cf7e994a99ed124bd3` |
| Train | 9,065 examples / 16 countries (cutoffs through Dec 2022) |
| Validation | 2,503 examples / 16 countries (Jan–Jul 2023 cutoffs) |
| Test | 2,230 examples / 14 countries (Aug–Dec 2023 cutoffs, target exactly +3 months) |
| Channels | 21 (six monthly observations each, plus observed-value masks) |
| Test label support | Phase 1: 854 · Phase 2: 976 · Phase 3: 395 · Phase 4: 5 · Phase 5: 0 |
| Annotation coverage | 2,781 / 9,065 training records (30.7%) — partial, not complete |

Prepared arrays live under `artifacts/dataset/` and `artifacts/multimodal/`
(`inputs.jsonl`, `targets.jsonl`, `manifest.json`). Loaders verify dataset checksums
against the version above before use.

### Sources

- **HFID** — district panel: FEWS NET IPC labels, normalized FCS/rCSI.
- **IMF PortWatch** — national shipping/port-call volumes.
- **ACLED** — national conflict events and fatalities.
- **CHIRPS** — country-average rainfall.
- **WFP** — national staple-food price series, with commodity/unit metadata.

National covariates are aggregated to country level; only FCS/rCSI and the IPC label
are at district (admin2) resolution. The benchmark assumes one-month release lags for
HFID/PortWatch/ACLED/WFP and two months for CHIRPS — it does not establish what could
actually have been forecast in real time. Full provenance:
[SOURCES.md](https://github.com/luk-huebner/OpenRelief/blob/main/docs/SOURCES.md),
[DATASET_CARD.md](https://github.com/luk-huebner/OpenRelief/blob/main/docs/DATASET_CARD.md).

## Annotations

`gpt-5.6-terra` generated structured training supervision for 2,781 of the 9,065
training examples: precursor claims (with channel citations), cross-domain hypotheses,
a written rationale, driver-cited recommended actions, and an uncertainty/confidence
note. **The teacher was given the true future IPC phase** — this is retrospective
supervision, not independent forecasting reasoning, and hypotheses are not proven
causes. Unannotated training examples train on an empty rationale/action target
(phase-only supervision).

Three fully worked examples — one deterioration, one persistent crisis, one
improvement, each with its input charts, generated hypotheses, verbatim rationale and
recommended actions — are in the
[annotation atlas](https://github.com/luk-huebner/OpenRelief/blob/main/docs/examples/annotation-atlas/README.md)
(exact examples: `docs/examples/annotation-atlas/examples.json`). It is an editorial
illustration of three selected cases, not a representative quality sample.

## Model results (this repo's `gpu-results/`)

`OpenTSLM/llama-3.2-1b-tsqa-sp` (Llama 3.2 1B backbone), fine-tuned with LoRA
(rank 16, alpha 32) on the 2,781 annotated training examples.

**Fine-tuning is not optional.** The untouched pretrained checkpoint, evaluated on the
full 2,230-example test set with no LoRA at all, produced **0% valid JSON output** —
it never attempts the task's output format. See `gpu-results/pretrained-fulltest/`.

**Source ablation** (256-example eval cohort, one source's channels dropped at a
time — see caveat below):

| Run | Macro-F1 | Accuracy |
|---|---:|---:|
| All sources (baseline) | 0.7199 | 0.7344 |
| Drop conflict (ACLED) | 0.7428 | 0.7617 |
| Drop rainfall (CHIRPS) | 0.7286 | 0.7500 |
| Drop prices (WFP) | 0.7300 | 0.7500 |
| **Drop shipping (PortWatch)** | **0.5826** | 0.7852 |

Dropping PortWatch (shipping) is the only ablation that meaningfully hurts the model —
a 0.14 macro-F1 drop, far outside the noise band of the other three. Full writeup:
[FINDING-portwatch-signal.md](https://github.com/luk-huebner/OpenRelief/blob/main/docs/FINDING-portwatch-signal.md).

*Caveat:* the ablation table above is a 256-example cohort with only one phase-4
example and no phase-5 examples in the full test set at all — do not read this as
demonstrated famine prediction. The clearer, full-test-set wins are the
pretrained-vs-fine-tuned gap above and the PortWatch ablation signal; match sample IDs
before comparing across cohorts. See
[STATUS.md](https://github.com/luk-huebner/OpenRelief/blob/main/docs/STATUS.md) for
the full, currently-known comparison table.

The fine-tuned LoRA adapter (`gpu-results/all-sources/best_model.pt`, SHA-256
`a10343caa152d1c3aa55b6dc9b40e11603067babeac44909001d1597a3e84e10`) is deployed as a
live inference endpoint that powers the project's demo frontend.

## Repository layout

```
artifacts/dataset/        prepared inputs.jsonl / targets.jsonl / manifest.json
artifacts/multimodal/     multimodal-formatted version of the same splits
artifacts/annotation-cache/  cached per-example teacher annotation responses
artifacts/*.manifest.json    acquisition manifests per source connector
gpu-results/               fine-tuning + ablation runs (see table above)
  all-sources/               baseline checkpoint, losses, benchmark, best_model.pt
  ablation-{acled,chirps,wfp,portwatch}/   one-source-dropped reruns
  pretrained-fulltest/        untouched-checkpoint baseline, full test set
  manifest.json                git commit + image digest + launch command per run
refrences/                 background reading + one raw source CSV
```

## Limitations

- National-level signals (shipping, conflict, rainfall, price) do not prove
  district-level exposure; only FCS/rCSI and the IPC label are district-resolution.
- FCS/rCSI normalization direction is undocumented — numeric movement is not given a
  severity interpretation.
- No phase-5 test examples exist and only five phase-4 examples — this dataset does
  not support a claim of demonstrated famine prediction.
- Ablation effects are associative, measured against removed input channels, not
  causal claims.
- Historical publication timestamps and revisions are unavailable; assumed release
  lags are a modeling choice, not an observed fact.

## License and reuse

Repository code is Apache-2.0. Access to an upstream source here does not imply
unrestricted redistribution rights over that source's original data — retain IMF
PortWatch terms, ACLED terms of use, WFP source metadata, CHIRPS attribution and
geoBoundaries attribution as applicable to each channel. See
[DATASET_CARD.md](https://github.com/luk-huebner/OpenRelief/blob/main/docs/DATASET_CARD.md#access-reuse-and-reproducibility)
for the full statement.