quirky chocolate origins

`ref` is a high-cardinality categorical with 630 distinct values across 2530 rows and no nulls, with entropy ratio 0.9954 indicating a near-uniform distribution. Values are short numeric strings (e.g. "414", "24", "1462") and the most frequent appears only 10 times (top_rate 0.0040), so this behaves like a reference/lookup id repeated a handful of times rather than a free-form feature. Treatment: treat as a foreign key and left-join to its reference table rather than one-hot encoding. high · anthropic:claude-opus-4-7

n

2,530

nulls

0 (0.0%)

unique

630

top_value

414

top_rate

0.003953

cardinality

630

entropy

9.257

entropy_ratio

0.9954

The column is labelled 'company' but contains a single value—an empty string—across all 2530 rows. Cardinality is 1, entropy is 0, and top_rate is 1.0, so it carries no information. This is effectively a placeholder field that was never populated. Treatment: Drop; constant empty value provides no signal. high · anthropic:claude-opus-4-7

n

2,530

nulls

0 (0.0%)

unique

1

top_value

top_rate

1

cardinality

1

entropy

0

entropy_ratio

0

Country of the chocolate maker, with 67 distinct locations and no nulls across 2530 rows. Heavily US-centric: 'U.S.A.' accounts for 44.9% (1136 rows), followed by Canada (177), France (176), and the U.K. (133), giving an entropy ratio of 0.61. Country names use abbreviated forms ('U.S.A.', 'U.K.') so any joins on canonical country lists will need normalisation. Treatment: Normalise country labels and group long-tail countries before one-hot or target encoding. high · anthropic:claude-opus-4-7

n

2,530

nulls

0 (0.0%)

unique

67

top_value

U.S.A.

top_rate

0.449

cardinality

67

entropy

3.675

entropy_ratio

0.6058

This column stores the year a review was recorded, ranging from 2006 to 2021 with only 16 unique values across 2530 rows. The distribution is centered around 2015 (mean 2014.37, median 2015) with a modest spread (std 3.97) and is roughly symmetric (skew -0.18). No nulls or outliers are present. Treatment: Treat as a year-level temporal feature; bin or convert to datetime for trend analysis. high · anthropic:claude-opus-4-7

n

2,530

nulls

0 (0.0%)

unique

16

min

2,006

max

2,021

mean

2014

median

2,015

std

3.968

q1

2,012

q3

2,018

iqr

6

skew

-0.1833

kurtosis

-0.7727

n_outliers

0

outlier_rate

0

zero_rate

0

Categorical country label identifying where the cocoa beans originated, with 62 distinct values across 2530 complete rows and no nulls. The distribution is broad rather than concentrated: the top value Venezuela accounts for only 10% of rows, and entropy ratio 0.79 confirms fairly even spread across many origins. Notable wrinkle: 'Blend' appears as the 6th most common value (156 rows), meaning some entries aren't a single country and will need special handling. Treatment: Group rare origins into 'Other', isolate 'Blend' as its own category, then one-hot or target-encode. high · anthropic:claude-opus-4-7

n

2,530

nulls

0 (0.0%)

unique

62

top_value

Venezuela

top_rate

0.1

cardinality

62

entropy

4.717

entropy_ratio

0.7921

This column captures the specific bean origin (region, estate, or country) for what appears to be a chocolate/cocoa dataset, with 1,605 unique values across 2,530 rows. Top values are dominated by countries like Madagascar (55), Ecuador (43), and Peru (41), but the high frequency of the word 'batch' (356 occurrences) suggests many entries mix origin names with batch identifiers, inflating uniqueness. Roughly 34% of values are single words and 37% are duplicates, indicating inconsistent granularity — some entries are broad countries, others are specific estates or batch-tagged labels. Treatment: Normalize by stripping batch suffixes and standardizing to country/region before encoding. high · anthropic:claude-opus-4-7

n

2,530

nulls

0 (0.0%)

unique

1,605

len_min

3

len_max

51

len_mean

17.12

len_median

14

len_p95

39

word_mean

2.681

word_median

2

n_empty

0

n_duplicates

925

duplicate_rate

0.3656

vocab_size

2,079

readability_flesch_mean

28.41

emoji_rate

0

url_rate

0

one_word_rate

0.3375

allcaps_rate

0.001581

boilerplate_rate

0

This is the cocoa percentage of each chocolate bar, ranging from 42 to 100 with a tight median of 70 and IQR of just 4. The distribution is right-skewed (skew 1.20, kurtosis 6.54) and 9.3% of rows flag as outliers — likely the high-cocoa tail pushing toward 100%. With only 46 unique values across 2530 rows, the field is effectively semi-discrete. Treatment: Use as-is or bin into cocoa-strength buckets; no transform needed given the narrow IQR. high · anthropic:claude-opus-4-7

n

2,530

nulls

0 (0.0%)

unique

46

min

42

max

100

mean

71.64

median

70

std

5.617

q1

70

q3

74

iqr

4

skew

1.198

kurtosis

6.541

n_outliers

235

outlier_rate

0.09289

zero_rate

0

This appears to be a coded recipe/composition field where each value lists a count followed by single-letter ingredient tokens (e.g. 'B,S,C' for what looks like beef/sauce/cheese-style components). With only 22 distinct combinations across 2530 rows and a top value ('3- B,S,C') covering 39.5% of records, the field is highly concentrated — entropy_ratio is just 0.545. Notably, 87 rows carry an empty string rather than null, so null_rate=0.0 understates true missingness. Treatment: Treat empty strings as missing, then one-hot encode the ingredient tokens (split on comma) rather than the raw combined string. high · anthropic:claude-opus-4-7

n

2,530

nulls

0 (0.0%)

unique

22

top_value

3- B,S,C

top_rate

0.3949

cardinality

22

entropy

2.43

entropy_ratio

0.545

Short free-text tasting notes (mean 23 characters, ~3 words) describing flavor and texture characteristics, almost certainly from a chocolate or cocoa review dataset given top tokens like 'cocoa', 'sweet', 'nutty', 'creamy', and 'fruit'. Values are near-unique (2487 distinct of 2530) yet built from a small vocabulary of 868 words, indicating these are comma-separated descriptor combinations rather than prose. Only 43 exact duplicates and no empties or URLs; readability mean of 49.7 is not very meaningful at this length. Treatment: Split on commas into descriptor tags and one-hot or embed for modelling. high · anthropic:claude-opus-4-7

n

2,530

nulls

0 (0.0%)

unique

2,487

len_min

3

len_max

37

len_mean

23.06

len_median

23

len_p95

30

word_mean

3.376

word_median

3

n_empty

0

n_duplicates

43

duplicate_rate

0.017

vocab_size

868

readability_flesch_mean

49.71

emoji_rate

0

url_rate

0

one_word_rate

0.007115

allcaps_rate

0

boilerplate_rate

0

A bounded numeric rating on a 1.0–4.0 scale with only 12 distinct values, suggesting half- or quarter-step increments rather than continuous scores. The distribution is tight (IQR 0.5, std 0.45) and slightly left-skewed (-0.61), centered near 3.25 with a mean of 3.20, and 50 low-end outliers (1.98%) pull the tail. No nulls or zeros, so every row carries a usable score. Treatment: Use as-is as an ordinal/numeric feature; consider treating the 50 low-end outliers separately if modelling tails. high · anthropic:claude-opus-4-7

n

2,530

nulls

0 (0.0%)

unique

12

min

1

max

4

mean

3.196

median

3.25

std

0.4453

q1

3

q3

3.5

iqr

0.5

skew

-0.6084

kurtosis

1.053

n_outliers

50

outlier_rate

0.01976

zero_rate

0