| --- |
| license: cc-by-nc-4.0 |
| tags: |
| - conflict-forecasting |
| - geospatial-transformer |
| - political-violence |
| - political-science |
| - ucdp |
| - pytorch |
| --- |
| |
| # UCDP Mini Spatial Event Encoder (ucdp_M) |
| |
| A small transformer encoder that learns representations of location-months from |
| UCDP Georeferenced Event Dataset (GED) event histories. Each PRIO-GRID 0.5 degree |
| cell-month is represented by its trailing 12-month sequence of organized-violence |
| events (type of violence, dyad sides, fatality bin, month offset), encoded with |
| masked-event modeling. The CLS embedding is a general-purpose representation for |
| conflict forecasting and transfer learning: fine-tune it for incidence forecasting, |
| rebel/state violence, escalation, or related event-based tasks. |
| |
| - Architecture: field embeddings -> concat-project -> [CLS] + TransformerEncoder |
| (4 layers, 8 heads, d=128), masked-event-modeling objective. |
| - Unit: PRIO-GRID 0.5 deg cell-month; max 200 events per sequence; seed 123. |
| - Training data: UCDP GED 25.1 (1989-2024), the open (CC BY 4.0) georeferenced |
| event dataset of organized violence. |
| |
| ## Usage |
| ```python |
| from modeling_ucdp import load_encoder |
| model, cfg = load_encoder(".") |
| # seq: LongTensor [B, 200, 5] (ids per ucdp_vocab.json; 0 = pad) |
| # mask: BoolTensor [B, 200] (True = real event) |
| z = model.encode_sequence(seq, mask) # [B, 128] cell-month embeddings |
| ``` |
| |
| ## License and attribution |
| Model weights and code: CC BY-NC 4.0. Training data is the UCDP GED (CC BY 4.0). |
| If you use this model, cite UCDP: |
| |
| - Davies, S., Pettersson, T., Sollenberg, M., & Oberg, M. (2025). Organized violence |
| 1989-2024, and the challenges of identifying civilian victims. Journal of Peace |
| Research 62(4). |
| - Sundberg, R. & Melander, E. (2013). Introducing the UCDP Georeferenced Event |
| Dataset. Journal of Peace Research 50(4). |
| |
| ## Limitations |
| - UCDP GED covers organized violence with at least one fatality; protests and |
| non-lethal events are not included. Coverage reflects media/source reporting. |
| - The encoder is a research artifact for representation learning and forecasting |
| research. It is not suitable for operational targeting or safety-critical use. |
| - Embeddings reflect historical reporting patterns and inherit their biases. |
| |
