Datasets:
File size: 2,307 Bytes
b9fd5cd | 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 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 | ---
license: other
task_categories:
- visual-question-answering
tags:
- benchability
- figure4
- vlm-sft
---
# BenchAbility Figure 4 -- `source_uniform`
One of two training mixtures drawn from the **same** frozen 884,143-row candidate pool, with the
**same** budget (60,000 intervention + 15,000 shared replay) and the same hyperparameters. The two
differ only in how the samples are chosen, which is the whole experiment.
| | |
|---|---|
| arm | `source_uniform` |
| selection | by source provenance only, chart:doc:ocr = 3:4:4 |
| intervention rows | 60,048 |
| replay rows | 15,000 |
| shards | 38 |
| pool | 884,143 rows / 20 source datasets, capability-tagged sample-by-sample |
## Columns
| column | meaning |
|---|---|
| `uid` | `source-split-index`, stable across both arms |
| `source` | original dataset (provenance) |
| `capability` | BenchAbility leaf, assigned per sample by a vision-language classifier |
| `split_role` | `intervention` or `replay` |
| `question` / `answer` | the training turn; `<image>` marks where the image goes |
| `image` | PNG bytes, embedded |
`capability` is present in **both** arms so the mixtures can be compared, but the
`source_uniform` draw never read it -- see below.
## How this arm was drawn
Sources are grouped into the three coarse families Figure 2 reports
(`chart`, `doc`, `ocr`) and drawn **3:4:4**. Within a family the quota is split across sources
proportional to sqrt(rows), then water-filled -- not equally, because equal shares would need FUNSD
(149 rows) roughly 11 times over while the chart family never repeated a row, and unequal repetition
between the arms would confound the comparison.
**This arm never reads a capability label.** The draw is handed rows with the field stripped. The arm
exists to model an engineer who has only benchmark-level reporting; letting it see sample-level
labels would make it a weaker copy of the other arm rather than the alternative it represents. Labels
are attached afterwards, for auditing what the draw happened to contain.
## Reproducing
```bash
python fig4_training/pipeline/40_mix.py # both arms from the frozen pool
python fig4_training/pipeline/50_export.py # this bundle
```
Full draw record, including every relaxed constraint and every shortfall, is in
`mixture_manifest.json`.
|