ValerianFourel/geodml-papersize
收藏资源简介:
--- license: cc-by-4.0 language: en pretty_name: GEODML — Generative-Engine Optimization Double Machine Learning size_categories: - 1M<n<10M tags: - llm - ranking - causal-inference - double-ml - rag - seo --- # GEODML — EMNLP 2026 submission End-to-end pipeline outputs for the GEODML paper. - 280 files, 5.3G - Generated: 2026-05-24T01:26:19+00:00 ## Layout ``` data/runs/ Stage A — LLM rerank outputs per (engine × model × pool × variant) data/order_probe/ Stage A' — order-sensitivity probe per seed data/features/ Stage B — engineered features (per engine × pool) data/main/ Stage C — main long-format experiment table per variant data/dml_results/ Stage D — DML treatment-effect estimates per variant interpretability/output/ Stage F — ablation / saliency / weights / probing CSVs docs/ paper-ready summaries (dml_summary_*, dml_headline.md) ``` ## Headline DML estimates See `docs/dml_headline.md` for the canonical PLR+LightGBM estimates, pooled across (engine × pool × model) subsets per (variant × treatment × outcome). ## Citation If you use this dataset, please cite the GEODML paper (EMNLP 2026).
GEODML (Generative-Engine Optimization Double Machine Learning) is a dataset submitted for EMNLP 2026, containing end-to-end pipeline outputs. This dataset focuses on generative engine optimization and double machine learning methods, covering stages such as LLM reranking, feature engineering, and treatment effect estimation. It includes 280 files with a total size of 5.3 GB, and was generated on May 24, 2026. Its structure encompasses multiple stages: Stage A (LLM reranking outputs), Stage A (sequential sensitivity probes), Stage B (engineered features), Stage C (main experimental tables), Stage D (DML treatment effect estimation), and interpretability outputs. This dataset is applicable to tasks including LLMs, ranking, causal inference, double machine learning, and RAG.




