Spaces:
Sleeping
Sleeping
Commit History
-
Merge value+attribute labels, add header label, RAG eval improvements
zofiasmolenasanaUnverified signatureLabel taxonomy: - Remove 'attribute' label, merge into 'value' (V key covers all data cells) - Add 'header' label (H key) for table titles/section headings - 'metadata' now only for footnotes, notes, disclaimers - Update UI legend, keyboard shortcuts, help guide (EN+PL) - Update LABELING_GUIDE.md for both languages - Fix all labeled data: attribute -> value (99k cells across 69 files) RAG evaluation: - Remove MAX_CELLS_FOR_PREDICTION skip from eval loop - Add naive chunking fallback for sheets too large for GNN - All 300 eval questions now evaluated (was losing 19% to skips) - Row-based chunking, aggregation enrichment, multiple model support New features: - RAG dashboard, eval pages, feature EDA - rag_baselines.py with pandas/beautifulsoup/lxml adapters - run_ablation_sweep.py for model comparison - layoutlm_encoder.py for LayoutLMv3 features - Extended model architectures and training pipeline Made-with: Cursord051588