saturn·

← all readings

vizwiz val annotations

saturn notebook · generated 2026-05-01 Report Notebook

Overview

Source: /home/coolhand/html/datavis/data_trove/cache/vizwiz_val_annotations.json

Saturn profiled 4,319 rows across 5 columns. The stats below are deterministic and machine-readable; the prose is a language-model interpretation of those stats (opt-in, added after the fact, never sees raw rows).

[2]:
!pip install saturn-dissect
import subprocess
subprocess.run([
    "saturn", "analyze", "/home/coolhand/html/datavis/data_trove/cache/vizwiz_val_annotations.json",
    "--findings", "vizwiz_val_annotations.json",
    "--llm", "anthropic:claude-opus-4-7",
])

Summary confidence: high

This dataset contains 4,319 rows from the VizWiz validation annotations, structured around image filenames, the questions asked about each image, the answers, an answer_type label, and an answerable flag. The questions column is the most interesting: about 35% are duplicates, with 'What is this?' alone appearing 523 times, suggesting a heavy concentration of generic identification queries. Answer_type is dominated by 'other' (62%) and 'unanswerable' (32%), and the answerable flag confirms that roughly 32% of items are flagged as not answerable — a key signal for any downstream modeling. The image column is uniquely identifying per row and not worth deeper analysis, while the answers column was skipped by the profiler.

citing: row_count · columns.question.stats.duplicate_rate · columns.question.top_values · columns.answer_type.top_values · columns.answer_type.stats.top_rate · columns.answerable.stats.mean · columns.answerable.stats.zero_rate · columns.question.language_counts

Out[4]:

saturn.schema() · 5 columns

column kind n null% unique alerts
image text 4,319 0.0% 4,319 near_unique one_word
question text 4,319 0.0% 2,798 multilingual duplicates
answers unknown 4,319 0.0% — skipped
answer_type categorical 4,319 0.0% 4
answerable numeric 4,319 0.0% 2
Fig 1.
answer_type · Shows the strong skew toward 'other' and 'unanswerable' categories that together cover ~94% of rows.
Show data table
valuecountshare
other269162.3%
unanswerable138532.1%
yes/no1954.5%
number481.1%
Fig 2.
answerable · Highlights that about 32% of questions are flagged as not answerable.
Show data table
bincount
0 – 0.0251385
0.025 – 0.050
0.05 – 0.0750
0.075 – 0.10
0.1 – 0.1250
0.125 – 0.150
0.15 – 0.1750
0.175 – 0.20
0.2 – 0.2250
0.225 – 0.250
0.25 – 0.2750
0.275 – 0.30
0.3 – 0.3250
0.325 – 0.350
0.35 – 0.3750
0.375 – 0.40
0.4 – 0.4250
0.425 – 0.450
0.45 – 0.4750
0.475 – 0.50
0.5 – 0.5250
0.525 – 0.550
0.55 – 0.5750
0.575 – 0.60
0.6 – 0.6250
0.625 – 0.650
0.65 – 0.6750
0.675 – 0.70
0.7 – 0.7250
0.725 – 0.750
0.75 – 0.7750
0.775 – 0.80
0.8 – 0.8250
0.825 – 0.850
0.85 – 0.8750
0.875 – 0.90
0.9 – 0.9250
0.925 – 0.950
0.95 – 0.9750
0.975 – 12934
Fig 3.
question · Top question values reveal how often generic prompts like 'What is this?' dominate the dataset.
Show data table
charscount
7 – 13759
13 – 20550
20 – 26931
26 – 33609
33 – 39368
39 – 46250
46 – 52143
52 – 58143
58 – 6596
65 – 7184
71 – 7868
78 – 8442
84 – 9135
91 – 9735
97 – 10330
103 – 11023
110 – 11619
116 – 12312
123 – 12923
129 – 13614
136 – 1428
142 – 14814
148 – 1558
155 – 1616
161 – 16810
168 – 1748
174 – 1808
180 – 1873
187 – 1935
193 – 2003
200 – 2063
206 – 2132
213 – 2191
219 – 2252
225 – 2321
232 – 2381
238 – 2450
245 – 2511
251 – 2580
258 – 2641
Fig 4.
question · Question length distribution shows most are short (median 26 chars) with a long tail up to 264 chars.
Show data table
charscount
7 – 13759
13 – 20550
20 – 26931
26 – 33609
33 – 39368
39 – 46250
46 – 52143
52 – 58143
58 – 6596
65 – 7184
71 – 7868
78 – 8442
84 – 9135
91 – 9735
97 – 10330
103 – 11023
110 – 11619
116 – 12312
123 – 12923
129 – 13614
136 – 1428
142 – 14814
148 – 1558
155 – 1616
161 – 16810
168 – 1748
174 – 1808
180 – 1873
187 – 1935
193 – 2003
200 – 2063
206 – 2132
213 – 2191
219 – 2252
225 – 2321
232 – 2381
238 – 2450
245 – 2511
251 – 2580
258 – 2641
Fig 5.
Per-column null rate across the corpus. Columns are ordered by input position.
Show data table
columnkindnull %
imagetext0.0%
questiontext0.0%
answersunknown0.0%
answer_typecategorical0.0%
answerablenumeric0.0%
Fig 6.
Language mix across all text columns (per-string detection, sampled).
Show data table
langcountshare
en430899.8%
la20.0%
es20.0%
hu10.0%
fy10.0%
ast10.0%
it10.0%
ia10.0%

image text identifier

This column holds image filenames following the pattern `vizwiz_val_########.jpg`, with all 4319 values being unique single tokens of exactly 23 characters. There are no nulls, duplicates, or vocabulary variation — every row maps one-to-one to a distinct image in what appears to be the VizWiz validation split. The negative Flesch score is an artifact of scoring filenames as prose and can be ignored.

Treatment: Use as a foreign key to load the corresponding image file; do not feed as text to a model.

anthropic:claude-opus-4-7 · confidence high
Out[12]:

saturn.columns["image"].stats

statvalue
n4,319
nulls0 (0.0%)
unique4,319
len_min 23
len_max 23
len_mean 23
len_median 23
len_p95 23
word_mean 1
word_median 1
n_empty 0
n_duplicates 0
duplicate_rate 0
vocab_size 4,319
readability_flesch_mean -47.98
emoji_rate 0
url_rate 0
one_word_rate 1
allcaps_rate 0
boilerplate_rate 0
alert: near_unique100.0% of rows are unique strings
alert: one_word100.0% rows are a single word
Fig 7.
Character-length distribution for image.
Show data table
charscount
22 – 230
23 – 230
23 – 230
23 – 230
23 – 230
23 – 230
23 – 230
23 – 230
23 – 230
23 – 230
23 – 230
23 – 230
23 – 230
23 – 230
23 – 230
23 – 230
23 – 230
23 – 230
23 – 230
23 – 230
23 – 234319
23 – 230
23 – 230
23 – 230
23 – 230
23 – 230
23 – 230
23 – 230
23 – 230
23 – 230
23 – 230
23 – 230
23 – 230
23 – 230
23 – 230
23 – 230
23 – 230
23 – 230
23 – 230
23 – 240

question text free_text

Short natural-language questions, mostly English (4308/4319) and overwhelmingly identification prompts — "What is this?" alone appears 523 times and the top 10 values are all generic "what is/color/says" queries. Heavy duplication (35.2%, 1521 rows) and a small vocab (2779 unique words across 4319 rows) suggest a VQA-style prompt set rather than diverse free text. Mean length is 35 chars / 7.3 words with very high Flesch readability (101.7), and a handful of non-English rows (es, la, it, fy, hu, ia, ast) introduce minor language drift.

Treatment: Tokenize and embed for modelling; deduplicate or weight by frequency given the 35% duplicate rate.

anthropic:claude-opus-4-7 · confidence high
Out[15]:

saturn.columns["question"].stats

statvalue
n4,319
nulls0 (0.0%)
unique2,798
len_min 7
len_max 264
len_mean 35.1
len_median 26
len_p95 95
word_mean 7.259
word_median 5
n_empty 0
n_duplicates 1,521
duplicate_rate 0.3522
vocab_size 2,779
readability_flesch_mean 101.7
emoji_rate 0
url_rate 0
one_word_rate 0
allcaps_rate 0.002547
boilerplate_rate 0.003473
alert: multilingual9 languages detected in sample
alert: duplicates35.2% duplicate strings
Fig 8.
Character-length distribution for question.
Show data table
charscount
7 – 13759
13 – 20550
20 – 26931
26 – 33609
33 – 39368
39 – 46250
46 – 52143
52 – 58143
58 – 6596
65 – 7184
71 – 7868
78 – 8442
84 – 9135
91 – 9735
97 – 10330
103 – 11023
110 – 11619
116 – 12312
123 – 12923
129 – 13614
136 – 1428
142 – 14814
148 – 1558
155 – 1616
161 – 16810
168 – 1748
174 – 1808
180 – 1873
187 – 1935
193 – 2003
200 – 2063
206 – 2132
213 – 2191
219 – 2252
225 – 2321
232 – 2381
238 – 2450
245 – 2511
251 – 2580
258 – 2641

answers unknown other

The column 'answers' was skipped by the profiler, so its kind is unknown and no descriptive statistics were computed. All 4319 rows are non-null, but uniqueness, type, and value distribution are unavailable. The name suggests it holds response content, likely structured (e.g., nested objects or arrays) which is why automatic profiling bailed out.

Treatment: Inspect raw values manually and parse into a typed structure before further profiling.

anthropic:claude-opus-4-7 · confidence low
Out[18]:

saturn.columns["answers"].stats

statvalue
n4,319
nulls0 (0.0%)
unique—
alert: skippedno profiler for kind=unknown

answer_type categorical label

Categorical label with only 4 distinct values across 4319 rows and no nulls, classifying answers into 'other', 'unanswerable', 'yes/no', and 'number'. The distribution is heavily imbalanced: 'other' covers 62.3% and 'unanswerable' another 1385 rows, while 'number' appears only 48 times. Entropy ratio of 0.61 confirms the skew toward the top two classes.

Treatment: One-hot or integer-encode; consider class-weighting or stratified sampling given the imbalance toward 'other'.

anthropic:claude-opus-4-7 · confidence high
Out[20]:

saturn.columns["answer_type"].stats

statvalue
n4,319
nulls0 (0.0%)
unique4
top_value other
top_rate 0.6231
cardinality 4
entropy 1.225
entropy_ratio 0.6127
Fig 9.
Top values for answer_type.
Show data table
valuecountshare
other269162.3%
unanswerable138532.1%
yes/no1954.5%
number481.1%

answerable numeric label

Binary 0/1 flag indicating whether a question is answerable, with 4319 rows and no nulls. Roughly 68% are marked answerable (mean 0.6793) and 32% are zeros, giving a moderate class imbalance toward the positive class. Only two unique values confirm this is a clean indicator rather than a probability score.

Treatment: Use directly as a binary target; account for the ~68/32 class imbalance during training or evaluation.

anthropic:claude-opus-4-7 · confidence high
Out[23]:

saturn.columns["answerable"].stats

statvalue
n4,319
nulls0 (0.0%)
unique2
min 0
max 1
mean 0.6793
median 1
std 0.4668
q1 0
q3 1
iqr 1
skew -0.7684
kurtosis -1.41
n_outliers 0
outlier_rate 0
zero_rate 0.3207
Fig 10.
Distribution of answerable. Vertical dash marks the median.
Show data table
bincount
0 – 0.0251385
0.025 – 0.050
0.05 – 0.0750
0.075 – 0.10
0.1 – 0.1250
0.125 – 0.150
0.15 – 0.1750
0.175 – 0.20
0.2 – 0.2250
0.225 – 0.250
0.25 – 0.2750
0.275 – 0.30
0.3 – 0.3250
0.325 – 0.350
0.35 – 0.3750
0.375 – 0.40
0.4 – 0.4250
0.425 – 0.450
0.45 – 0.4750
0.475 – 0.50
0.5 – 0.5250
0.525 – 0.550
0.55 – 0.5750
0.575 – 0.60
0.6 – 0.6250
0.625 – 0.650
0.65 – 0.6750
0.675 – 0.70
0.7 – 0.7250
0.725 – 0.750
0.75 – 0.7750
0.775 – 0.80
0.8 – 0.8250
0.825 – 0.850
0.85 – 0.8750
0.875 – 0.90
0.9 – 0.9250
0.925 – 0.950
0.95 – 0.9750
0.975 – 12934

How to cite

click to copy

BibTeX
@misc{saturn-vizwiz-val-annotations-2026,
  author       = {Steuber, Luke},
  title        = {Saturn reading: vizwiz val annotations},
  year         ={2026},
  howpublished = {\url{https://dr.eamer.dev/saturn/view/vizwiz_val_annotations}},
  note         = {Profiled with saturn-dissect v0.2.0, prompt saturn-insight-v2, model anthropic:claude-opus-4-7},
}
APA
Steuber, L. (2026). Saturn reading: vizwiz val annotations. Source: /home/coolhand/html/datavis/data_trove/cache/vizwiz_val_annotations.json. Profiled with saturn-dissect v0.2.0 (saturn-insight-v2, anthropic:claude-opus-4-7). Retrieved from https://dr.eamer.dev/saturn/view/vizwiz_val_annotations