README.md

---
pretty_name: VectorOS Vector 100k SimSat VLM Dataset
license: other
language:
- en
task_categories:
- visual-question-answering
- text-generation
tags:
- geospatial
- remote-sensing
- public-health
- vector-borne-disease
- sentinel-2
- mapbox
- multimodal
- lfm2-vl
- lfm25-vl
- weak-supervision
configs:
- config_name: records
  data_files:
  - split: train
    path: data/train.jsonl
  - split: validation
    path: data/validation.jsonl
  - split: test
    path: data/test.jsonl
---



# VectorOS Vector 100k SimSat VLM Dataset

VectorOS Vector 100k is a high-fidelity multimodal instruction dataset for fine-tuning vision-language models on geospatial epidemiology tasks. It was built for the VectorOS hackathon project and targets `LiquidAI/LFM2.5-VL-450M`.

The dataset contains **100,000** chat-style examples derived from **10,000** geospatial chips across **30 AOIs**. Every accepted chip has a real SimSat Sentinel-2 true-color view, a real SimSat Sentinel-2 NIR-red-green false-color view, a real Mapbox satellite view, and an aligned open-layer evidence overlay.

Created for upload: `2026-05-06T17:28:22Z`

![dataset_montage](https://cdn-uploads.huggingface.co/production/uploads/6375af60e3413701a9f01c0f/ZktzLpUrpBiI5clDaf1ah.png)

## What Is Included

Each chip includes:

- `image_packets/<aoi>/<chip>_packet.png`: a 1024 x 1024 four-panel visual packet.
- `sidecars/<aoi>/<chip>_sidecar.json`: numeric features, source paths, provenance pointers, quality fields, and license flags.
- `targets/<aoi>/<chip>_risk_tile.json`: strict VectorOS risk-tile target JSON.
- `raw_simsat/<aoi>.tar`: per-AOI tar shard containing the raw per-chip SimSat products:
  - `sentinel_rgb.png`
  - `sentinel_false_color_nir_red_green.png`
  - `sentinel_bands_red_green_blue_nir.npz`
  - `sentinel_metadata.json`
  - `mapbox_satellite.png`
  - `mapbox_metadata.json`

The four-panel image packet order is:

1. top-left: SimSat Sentinel-2 true-color RGB
2. top-right: SimSat Sentinel-2 false color NIR-red-green
3. bottom-left: Mapbox satellite context
4. bottom-right: aligned evidence overlay from ESA WorldCover, JRC Global Surface Water, CHIRPS rainfall, WorldPop, OSM, and weak vector/disease labels

## Loadable Splits

The `records` config exposes the VLM instruction records:

| Split | Examples |
|---|---:|
| train | 72,000 |
| validation | 14,000 |
| test | 14,000 |

```python
from datasets import load_dataset

ds = load_dataset("Alfaxad/vector-100k", "records")
print(ds)
print(ds["train"][0]["messages"])
```

The `image`, `sidecar`, and `target` fields are repository-relative paths. For local fine-tuning, download the repository snapshot and resolve these paths against the snapshot root:

```python
from pathlib import Path
from huggingface_hub import snapshot_download
from datasets import load_dataset

root = Path(snapshot_download(repo_id="Alfaxad/vector-100k", repo_type="dataset"))
ds = load_dataset("Alfaxad/vector-100k", "records")
row = ds["train"][0]
packet_path = root / row["image"]
sidecar_path = root / row["sidecar"]
target_path = root / row["target"]
```

To extract raw SimSat assets for a specific AOI:

```python
import tarfile
from pathlib import Path
from huggingface_hub import snapshot_download

root = Path(snapshot_download(repo_id="Alfaxad/vector-100k", repo_type="dataset"))
with tarfile.open(root / "raw_simsat" / "bangkok_dengue.tar") as tar:
    tar.extractall(root / "raw_simsat_extracted" / "bangkok_dengue")
```

## Tasks

Each chip contributes 10 task variants:

- `risk_tile_json`
- `officer_explanation`
- `evidence_cards_json`
- `uncertainty_audit_json`
- `field_task_brief`
- `habitat_patch_summary`
- `exposure_summary_json`
- `hard_negative_assessment`
- `source_provenance_json`
- `copilot_why_here`

Every task is grounded in the same image packet and sidecar. All task variants from a chip stay in the same split.

## AOIs And Modules

The dataset covers three initial VectorOS disease modules:

- Dengue/Aedes: Bangkok, Cairns, Colombo, Dhaka, Iquitos, Panama City, Recife, Rio de Janeiro, San Juan, Singapore.
- Malaria/Anopheles: Bobo-Dioulasso, Farafenni, Ifakara, Kisumu, Korhogo, Manhica, Navrongo, Nchelenge, Siaya, Tororo.
- Schistosomiasis: Busia, Jinja, Kafr el-Sheikh, Kafue, Mangochi, Mbita/Homa Bay, Mwanza, Mwea, Niamey, Richard Toll.

Module chip counts:

- dengue/Aedes: 3,334
- malaria/Anopheles: 3,333
- schistosomiasis: 3,333

## Sampling Strategy

The chips were stratified to train useful visual and reasoning behavior rather than only label memorization:

- label-positive jitter around GBIF, MAP, or OpenDengue evidence
- hard negatives near water or urban context with no weak labels
- exposure-context chips around populated or operationally relevant areas
- random ecological negatives
- uncertain sparse-context chips

Sample type counts:

- label-positive jitter: 3,524
- hard negative water/urban: 2,762
- exposure context: 1,568
- random ecological: 1,472
- uncertain sparse context: 674

## Source Layers

The dataset combines:

- SimSat Sentinel-2 products through Earth Search / Sentinel-2 L2A
- Mapbox satellite static imagery
- CHIRPS v2 daily rainfall
- JRC Global Surface Water v1.4
- ESA WorldCover 2021 v200
- WorldPop population surfaces
- OpenStreetMap / OSM healthcare and operational context
- OpenDengue V1.3 where applicable
- Malaria Atlas Project extracts for malaria AOIs
- GBIF occurrence records for vector or intermediate-host evidence

## Validation

Build-time validation:

- chips: 10,000
- total examples: 100,000
- packets exist: 10,000
- Sentinel packets available: 10,000
- Mapbox packets available: 10,000
- target schemas valid: 10,000
- VLM record schemas valid: 100,000

Additional local validation before upload:

- 40,000 referenced images were header/size verified.
- A 400-image random variance spot check found zero low-variance samples.
- The staged package was scanned for temporary HF/Mapbox token leakage before upload.
- The public-release Mapbox flag is `user_verified_for_project_use`.

## Safety And Intended Use

This dataset is intended for fine-tuning and evaluating geospatial VLM behavior for public-health surveillance support. It should be used for population-level, decision-support style reasoning only.

The prompts and targets explicitly avoid:

- individual health claims
- claims of field-verified local disease presence
- individual care or treatment guidance
- calibrated epidemiological forecasting claims

## Limitations

- Labels are weak supervision, not ground truth.
- GBIF, MAP, OpenDengue, and other open surveillance records are presence/survey-biased.
- Absence of weak labels is not evidence of field absence.
- Risk scores are instruction-tuning targets generated from open environmental, exposure, and label proxies, not calibrated predictions.
- Mapbox redistribution is marked under the project-specific user verification flag.

## Repository Layout

```text
README.md
data/
  train.jsonl
  validation.jsonl
  test.jsonl
image_packets/
  <aoi>/*.png
sidecars/
  <aoi>/*.json
targets/
  <aoi>/*.json
raw_simsat/
  <aoi>.tar
metadata/
  chip_index.json
  chip_index.csv
  splits.json
  manifest.json
  provenance.json
  validation_summary.json
  raw_simsat_index.json
  hf_layout.json
schemas/
  risk_tile_target.schema.json
  vlm_record.schema.json
```

## Citation

If you use this dataset, cite it as:

```bibtex
@dataset{vectoros_vector_100k_simsat,
  title = {Vector 100k SimSat VLM Dataset},
  author = {Alfaxad Eyembe},
  year = {2026},
  publisher = {Hugging Face},
  url = {https://huggingface.co/datasets/Alfaxad/vector-100k}
}
```