Datasets:

mlsys-io
/

EventXBench

data/t3/train.jsonl (279,924 rows) replaces the old test.jsonl (342,552 rows, stale) - rebuilt from T3_Reproducible_Package's silver export (t3_silver_labels_full.parquet) joined with deterministic-check results, LLM grading output, raw tweet/market text, and market predicates. Renamed test -> train since no code in this repo treats it as held-out data; baselines self-split it 70/30 by condition_id at runtime.
data/t3/gold.jsonl (2,687 rows) is new: the human-adjudicated gold audit pool (gold_grade, resolved via majority vote / pair agreement / senior adjudication), joined to raw tweet/question/resolution-rule text so it can be queried directly for LLM grading without going through the (GPT-derived, potentially circular) predicate field.
The old, unused train.jsonl (25.4MB, never referenced by any loader code) is overwritten by the new one above at the same path.

Both files verified end-to-end against the updated eventxbench loader and baseline scripts in mlsys-io/EventXBench#8 - see that PR for the corresponding code changes (gold-aware evaluation, LightGBM feature-leakage fix reproducing the package's exact 0.4562 kappa / 0.3963 macro-F1, etc.).

tjluyao changed pull request status to merged 8 days ago

Upload images, audio, and videos by dragging in the text input, pasting, or clicking here.

Tap or paste here to upload images

· Sign up or log in to comment