Upload README.md with huggingface_hub

README.md

@@ -0,0 +1,92 @@
+---
+license: cc-by-4.0
+language:
+- en
+pretty_name: GDELT-Forecast Binary
+size_categories:
+- 1K<n<10K
+task_categories:
+- text-generation
+- question-answering
+tags:
+- forecasting
+- temporal-reasoning
+- gdelt
+- news
+- evaluation
+---
+
+# GDELT-Forecast Binary
+
+1,215 yes/no forecasting questions generated from clusters of news articles in the GDELT 2.0 corpus (Aug 2025 – Apr 2026). Each question is paired with the original seed-event articles, top-5 retrieved evidence articles dated strictly before the question creation date, and a verified ground-truth answer.
+
+## Intended use
+
+Training and evaluating LLM-based forecasting models in a *strict forecasting posture* — the model sees only news that was publicly available before the question's creation date, then predicts the resolution. Companion to [`gdelt-forecast-freeform`](https://huggingface.co/datasets/rajatagarwal457/gdelt-forecast-freeform), which contains 924 non-binary questions (name, number, date, free-form) built from the same corpus.
+
+## How it was built
+
+Five-stage pipeline over the GDELT 2.0 corpus (~31.7 M articles):
+
+1. **Politics/geopolitics filter** — GPT-4o-mini classifies articles by topic; ~2 % retained.
+2. **Event clustering** — articles describing the same event clustered with a similarity-graph approach. 2,721 clusters of size 3-5 retained.
+3. **Question generation** — GPT-4o reads each cluster and writes a forecasting question that resolves *after* the most recent article in the cluster. Answer type pre-assigned (here: yes/no). Sources must agree on the answer.
+4. **Evidence retrieval** — for each question, top-5 articles retrieved from the corpus via TF-IDF ∪ OpenAI-embed FAISS → OpenAI rerank, filtered to publication date strictly < question creation date.
+5. **Aggregation** — questions with ≥ 5 evidence articles kept. No similarity-based dedup at this stage (we leave that to downstream consumers).
+
+## Schema
+
+| Column | Type | Description |
+|---|---|---|
+| `qid` | int | Synthetic question ID. Stable within this release. |
+| `cluster_id` | int | ID of the source event cluster. |
+| `cluster_size` | int | Number of articles in the source cluster (3-5). |
+| `question_title` | string | The forecasting question itself. |
+| `background` | string | One- or two-sentence summary of the event context. |
+| `resolution_criteria` | string | How the answer is defined and what counts as a match. |
+| `answer_type` | string | Always `yes_no` in this dataset. |
+| `ground_truth_answer` | string | Either `yes` or `no`. |
+| `question_start_date` | string (YYYY-MM-DD) | The cutoff. **No information from this date or later is allowed when forecasting.** |
+| `resolution_date` | string (YYYY-MM-DD) | When the question's outcome is known. |
+| `seed_urls` | list[string] | URLs of the original cluster articles (the ones GPT-4o read to generate the question). |
+| `seed_dates` | list[string] | Publication dates of the seed articles. |
+| `evidence_articles` | list[struct] | Top-5 retrieved evidence articles. Each: `{url, date, source, text}`. All dated strictly before `question_start_date`. |
+| `sources_agree` | bool | Whether the seed articles' implied answers agreed during generation. 5 rows have `False`; treat with caution. |
+| `quality_flags` | list[string] | Empty in this release. |
+| `required_type` | string | The answer type that was requested at generation time (always `yes_no` here). |
+
+## Statistics
+
+- **Total questions**: 1,215
+- **Cluster size**: 3 → 538, 4 → 251, 5 → 426
+- **Sources agree**: 1,212 / 1,215
+- **Date range** (`question_start_date`): late Aug 2025 → mid Mar 2026
+- **Median forecast horizon** (cutoff → resolution): ~14 days
+- **Ground-truth balance**: roughly 60 / 40 yes / no
+
+## Strict forecasting posture
+
+`question_start_date` is the cutoff. The retrieval pipeline filters all evidence to dates `< question_start_date`. If you build your own retrieval, **respect this cutoff** — articles dated on or after `question_start_date` may contain explicit answers, leading to leakage.
+
+## Known limitations
+
+- **GDELT date metadata** is trusted; we have not audited mis-dating systematically.
+- Some questions may be **answerable from priors alone** if the model has memorised related real-world events. We recommend pairing with a no-context baseline run.
+- Some questions are **near-duplicates** of others (no dedup at this stage). If you need a dedup'd subset, embed `question_title` and drop pairs with cosine sim ≥ 0.85.
+
+## Citation
+
+If you use this dataset, please cite:
+
+```bibtex
+@misc{anthral2026gdeltforecast,
+  title  = {GDELT-Forecast: Strict-cutoff forecasting benchmarks from GDELT news clusters},
+  author = {Rajat Agarwal and Anthral Labs},
+  year   = {2026},
+  url    = {https://huggingface.co/datasets/rajatagarwal457/gdelt-forecast-binary},
+}
+```
+
+## License
+
+CC-BY-4.0, matching the upstream GDELT 2.0 license. Attribution: cite this dataset and acknowledge GDELT as the source corpus.