File size: 1,550 Bytes
02ec006
 
 
 
 
 
cd1a734
02ec006
 
cd1a734
02ec006
cd1a734
 
 
02ec006
 
 
 
 
 
 
 
 
cd1a734
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
---
license: other
task_categories:
- visual-question-answering
tags:
- benchability
- figure4
---

# BenchAbility Figure 4 -- held-out eval

The 10% candidate-dev half of the same 884,143-row pool the two training mixtures are drawn from,
balanced per capability. Split **by image**, so no picture here appears in either mixture, and both
arms are equally blind to it.

| capability | n |
|---|---:|
| chart_reading / chart_reasoning | 250 / 250 |
| table_lookup / table_reasoning | 250 / 250 |
| document_qa / document_text_reading | 250 / 250 |
| diagram_and_infographic_understanding | 250 |
| scene_text_recognition | 250 |
| key_information_extraction | 168 (all that exists) |

Balanced rather than proportional: the pool is 38% `chart_reasoning`, so a proportional dev set would
measure that leaf precisely and the scarce ones not at all -- and the scarce ones are where the two
arms differ most. One row per image, because a second question on the same picture is not an
independent measurement.

**This is an in-domain dev set, not the paper's evaluation.** Same 20 sources, held out by image. It
answers "did training move this capability at all", which separates a broken run from a real one. It
cannot separate "learned the capability" from "learned these 20 datasets" -- that needs the
unseen-source set (ChartQAPro, OCRBench v2, DUDE, MME-RealWorld) which is not in this pool.

Score with `evaluate.py` from the training bundle; measure the base model first, then each
checkpoint, and report `gain = checkpoint - base` per capability.