File size: 2,805 Bytes
3a2675d
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
63
64
65
66
67
68
69
70
71
72
---
license: other
task_categories:
- visual-question-answering
tags:
- benchability
- figure4
- vlm-sft
---

# BenchAbility Figure 4 -- `capability_guided`

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 | `capability_guided` |
| selection | by capability, gap-weighted from the Figure 3 scores |
| intervention rows | 59,999 |
| 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

Per-capability quotas come from the measured Figure 3 scores:

    gap(l)    = max(0, 95 - score(l))
    weight(l) = 0.20/9 + 0.80 * gap(l) / sum(gap)

The scores are measured on the diagnosis benchmarks through `scripts/build_fig1.py`'s own code path,
not the placeholder table in the plan document. The target is 95 rather than 70 because the real
scores top out at 93.69 -- at 70, seven of the nine capabilities have gap 0 and the experiment
collapses into "train two capabilities".

Within each capability bucket, samples are drawn **across benchmarks** under three diversity bounds:
a 50% per-source cap, a 15% per-template cap, and at most 2 rows per image. Where a bound is
arithmetically unsatisfiable it is relaxed one at a time and the relaxation is recorded --
`key_information_extraction` is three templates in total by construction, and AI2D is 71% of
`diagram_and_infographic_understanding`.

Two capabilities cannot be filled from this pool: diagram (7,264 available against 11,081 requested)
and KIE (1,416 against 1,864). The deficit is redistributed over the capabilities that have room, in
gap proportion. It is **never** filled by upsampling with replacement, which would give those buckets
more effective epochs than the same rows get in the other arm.

## 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`.