TABVERSE logo

Leaderboard

QA  ·  SUC  ·  SR  ·  17 Models  ·  3 Formats  ·  3 Pipelines

← Back to Paper Page PDF Paper GitHub Code Hugging Face Dataset
Filter:
High
Mid
Low
Very Low

Strict exact-match accuracy (%). Click any column header to sort. VLM-Image columns show image pipeline; Text columns show VLM-Text / LLM-Text pipeline.

Model Img Avg Img HTML Img LaTeX Img Md Text Avg Text HTML Text LaTeX Text Md

Table 3 from TABVERSE paper. Bold = best per column. * = limited context window; scores over available instances.

Pipeline:
Format:

Exact-match accuracy (%) on 10 structure-oriented subtasks. T.P.=Table Partition, F.C.=First Cell, L.C.=Last Cell, S.D.=Size Detection, C.Lu.=Cell Lookup, R.Lu.=Reverse Lookup, Co.Rt.=Column Retrieval, Ro.Rt.=Row Retrieval.

Model Overall T.P. F.C. L.C. S.D. #Rows #Cols C.Lu. R.Lu. Co.Rt. Ro.Rt.

Tables 4, 12, 17 from TABVERSE paper. Select pipeline and format above.

Source format:
≥0.95
0.85–0.95
0.70–0.85
<0.70

GriTS-Topology (structural fidelity) and GriTS-Content (cell-text fidelity) for three target formats, plus syntactic usability fraction. Higher = better.

Model → HTML target → Markdown target → LaTeX target
Topology Content Usability Topology Content Usability Topology Content Usability

Tables 7 & 8 from TABVERSE paper. GriTS scores in [0,1]; usability = fraction of syntactically valid outputs.