QA report: emotion

The data-quality report on this adapter's training, development, calibration and test files, written by the maintainers' QA script before training and reviewed by someone who did not build the data. For publication, internal file paths were cut to file names and machine, service and account names were removed; every number, verdict and sample row is unchanged.

QA: emotion

Checked 2026-09-30 11:12 by adapters/qa/qa.py (READY file ready-emotion.json, public v1 (READY 2026-09-29T23:56)).

Verdict: PASS WITH NOTES (for v2)

Findings

No shortcut. A surface-only model sits at the majority baseline (31.1% accuracy, 5% balanced against 3.6% chance). The strong phrases ("sorry" for remorse, "thanks" for gratitude, "lol" for amusement) are the meaning of the labels. The GoEmotions placeholders "[NAME]" and "[RELIGION]" appear in 13.9% of rows, spread evenly across classes (9–17%).
Label noise is built into the source. GoEmotions is multi-label, and v1 keeps one label per comment. Near-identical comments carry different labels: "I'm sorry for your loss" is labelled caring, grief, remorse and amusement. The BoW model's balanced accuracy is only 32%, which fits a noisy task. v2: turn the raters' labels into soft targets (the counts are in source.gold_labels and upstream_ids), and/or merge rare, confusable classes.
The class balance is heavily skewed: neutral is 31% of test. Duplicates are negligible (0.1%), and held-out near duplicates are 0.1–0.2%.

Automatic flags (for the reviewer to judge; not all are problems)

formatting 'ends with ?' differs by class: curiosity 51%, confusion 32%, neutral 5%, surprise 5% …
formatting 'ends with .' differs by class: disapproval 62%, realization 58%, approval 57%, disappointment 56% …
formatting 'ends with !' differs by class: excitement 29%, gratitude 23%, joy 15%, admiration 15% …
formatting 'no end punctuation' differs by class: amusement 54%, grief 43%, sadness 41%, love 40% …
157 strong phrase flags (see list): review whether they are meaning or leakage

Data checked

split	rows	families	file
train	46,067	46067	`train.jsonl`
dev	1,000	1000	`dev.jsonl`
calibration	1,000	1000	`calibration.jsonl`
test	5,408	5408	`test.jsonl`

Train sha256: 30867b255a4e4ca1eb9ec1f3116245aebacd6cf702a0a9adc0e71c87cd1c4bcd (READY file gives no checksum)
Main text field (the text the phrase and length checks use): state.comment.
Label classes: admiration, amusement, anger, annoyance, approval, caring, confusion, curiosity, desire, disappointment, disapproval, disgust, embarrassment, excitement, fear, gratitude, grief, joy, love, nervousness, neutral, optimism, pride, realization, relief, remorse, sadness, surprise (<listed option> = one of the per-row listed options such as t3 or o12). Row kinds (source.kind): emotion.

3. Balance

Label class share per split

lclass	train	dev	calibration	test	train rows
admiration	7.8%	7.4%	7.7%	7.8%	3,590
amusement	4.6%	5.7%	5.4%	4.2%	2,102
anger	2.9%	3.0%	2.9%	3.1%	1,357
annoyance	4.5%	3.8%	3.7%	4.8%	2,082
approval	5.6%	5.7%	7.6%	5.2%	2,566
caring	2.0%	2.5%	2.0%	2.0%	910
confusion	2.6%	2.5%	1.7%	2.4%	1,180
curiosity	4.1%	2.8%	5.4%	4.2%	1,872
desire	1.2%	0.5%	1.7%	1.3%	542
disappointment	2.3%	2.3%	1.8%	2.0%	1,061
disapproval	3.9%	4.6%	5.3%	4.3%	1,818
disgust	1.5%	1.3%	1.2%	1.8%	675
embarrassment	0.6%	0.6%	0.4%	0.6%	260
excitement	1.5%	0.9%	1.6%	1.4%	693
fear	1.2%	1.0%	1.1%	1.3%	570
gratitude	4.9%	6.0%	5.3%	5.4%	2,279
grief	0.1%	0.2%	0.0%	0.1%	67
joy	2.6%	2.2%	2.7%	2.4%	1,188
love	3.9%	3.5%	3.4%	3.7%	1,808
nervousness	0.3%	0.2%	0.4%	0.3%	130
neutral	31.3%	33.0%	28.1%	31.1%	14,417
optimism	2.8%	2.8%	3.5%	2.7%	1,293
pride	0.2%	0.4%	0.3%	0.2%	81
realization	1.9%	2.8%	2.0%	2.3%	865
relief	0.3%	0.0%	0.3%	0.1%	126
remorse	1.0%	0.7%	0.8%	0.9%	474
sadness	2.5%	1.3%	2.0%	2.4%	1,129
surprise	2.0%	2.3%	1.7%	2.0%	932

Row kind share per split

kind	train	dev	calibration	test	train rows
emotion	100.0%	100.0%	100.0%	100.0%	46,067

4. Format

split	row-level format problems
train	none
dev	none
calibration	none
test	none

Options per choice row

split	min	median	p99	max
train	28	28	28	28
dev	28	28	28	28
calibration	28	28	28	28
test	28	28	28	28

Prompt length in tokens

split	measure	median	p99	max
train	estimate: characters / 3 (upper bound for English)	669	701	870
dev	estimate: characters / 3 (upper bound for English)	669	705	710
calibration	estimate: characters / 3 (upper bound for English)	669	701	708
test	estimate: characters / 3 (upper bound for English)	668	700	713

State key sets (train)

keys	rows
comment, source	46,067 (100.0%)

Instructions (train)

Canonical (the most common text) 70.1%, reworded 26.9% (20 distinct rewordings), none 3.0%. Target about 70 / 27 / 3.
Canonical text: "Which emotion does the comment express most? Choose the emotion that is strongest in the comment, or neutral if it expresses no particular emotion."
source instruction tag: canonical 70.1%, none 3.0%, variant-18 1.4%, variant-14 1.4%, variant-12 1.4%, variant-19 1.4%

class	canonical	none
admiration	70.1%	2.9%
amusement	69.0%	2.8%
anger	68.3%	3.8%
annoyance	69.5%	3.1%
approval	70.6%	3.5%
caring	72.0%	1.9%
confusion	71.9%	2.6%
curiosity	68.9%	2.6%
desire	71.0%	2.6%
disappointment	69.5%	3.6%
disapproval	71.6%	2.6%
disgust	72.1%	3.0%
embarrassment	71.9%	2.7%
excitement	69.0%	4.2%
fear	69.5%	3.9%
gratitude	69.1%	2.6%
grief	68.7%	4.5%
joy	70.2%	2.9%
love	70.2%	3.5%
nervousness	66.2%	4.6%
neutral	70.0%	3.0%
optimism	71.0%	3.4%
pride	72.8%	3.7%
realization	70.1%	3.8%
relief	67.5%	4.8%
remorse	66.9%	3.6%
sadness	72.7%	2.5%
surprise	72.2%	2.7%

1. Shortcuts

Phrase statistics and models use a label-stratified sample of 46,067 train rows; models are scored on the full test file (5,408 rows).

Text length by label class (main text, characters)

split	class	rows	p10	median	p90	mean
train	admiration	3590	20	58	117	63
train	amusement	2102	22	63	113	65
train	anger	1357	18	58	121	64
train	annoyance	2082	28	73	126	75
train	approval	2566	26	71	121	72
train	caring	910	25	72	122	73
train	confusion	1180	31	76	121	75
train	curiosity	1872	27	66	120	70
train	desire	542	31	69	118	70
train	disappointment	1061	33	74	124	76
train	disapproval	1818	30	73	124	75
train	disgust	675	25	67	118	69
train	embarrassment	260	33	74	122	76
train	excitement	693	20	55	111	60
train	fear	570	24	69	121	72
train	gratitude	2279	22	59	114	64
train	grief	67	25	57	114	64
train	joy	1188	23	64	116	67
train	love	1808	20	60	115	64
train	nervousness	130	25	77	124	75
train	neutral	14417	20	64	121	67
train	optimism	1293	36	80	124	80
train	pride	81	26	54	111	61
train	realization	865	36	77	126	79
train	relief	126	27	64	109	66
train	remorse	474	29	69	122	72
train	sadness	1129	22	69	119	69
train	surprise	932	25	65	118	69
test	admiration	422	19	60	110	62
test	amusement	227	24	61	111	64
test	anger	166	18	63	124	67
test	annoyance	261	26	73	126	75
test	approval	281	25	70	123	72
test	caring	110	22	68	125	71
test	confusion	128	32	76	118	75
test	curiosity	225	22	60	112	63
test	desire	69	38	78	114	73
test	disappointment	109	32	71	127	76
test	disapproval	232	28	72	122	74
test	disgust	96	26	67	123	72
test	embarrassment	30	31	88	128	79
test	excitement	78	20	50	117	61
test	fear	72	25	63	118	68
test	gratitude	294	19	58	112	61
test	grief	4	36	75	82	64
test	joy	128	23	63	115	67
test	love	198	20	60	112	62
test	nervousness	17	35	71	126	76
test	neutral	1682	19	64	119	66
test	optimism	146	31	75	121	74
test	pride	13	35	41	93	55
test	realization	124	36	83	121	79
test	relief	8	32	59	92	61
test	remorse	51	27	56	104	61
test	sadness	129	18	66	124	68
test	surprise	108	22	61	113	64

By row kind (train): main-text length, length of the rest of the state, options

kind	rows	median chars	mean chars	median other-state chars	median options
emotion	46067	66	69	61	28

Correct option: longest / shortest / position / key

For rows whose answer is one of the listed options (fixed options such as 'none of these' excluded):

split	rows	correct is longest	correct is shortest	chance (1/listed)	mean relative position (0 first, 1 last; 0.5 expected)	position fifths

Correct key and position by option count

split	options	rows	mean options	top correct keys	most common position (0-based)
train	11-30	46067	28.0	neutral 31.3%, admiration 7.8%, approval 5.6%, gratitude 4.9%, amusement 4.6%	2 (4.2%)
test	11-30	5408	28.0	neutral 31.1%, admiration 7.8%, gratitude 5.4%, approval 5.2%, annoyance 4.8%	1 (4.5%)

Option count by label class (train)

class	rows	min	median	mean	max
admiration	3590	28	28	28.0	28
amusement	2102	28	28	28.0	28
anger	1357	28	28	28.0	28
annoyance	2082	28	28	28.0	28
approval	2566	28	28	28.0	28
caring	910	28	28	28.0	28
confusion	1180	28	28	28.0	28
curiosity	1872	28	28	28.0	28
desire	542	28	28	28.0	28
disappointment	1061	28	28	28.0	28
disapproval	1818	28	28	28.0	28
disgust	675	28	28	28.0	28
embarrassment	260	28	28	28.0	28
excitement	693	28	28	28.0	28
fear	570	28	28	28.0	28
gratitude	2279	28	28	28.0	28
grief	67	28	28	28.0	28
joy	1188	28	28	28.0	28
love	1808	28	28	28.0	28
nervousness	130	28	28	28.0	28
neutral	14417	28	28	28.0	28
optimism	1293	28	28	28.0	28
pride	81	28	28	28.0	28
realization	865	28	28	28.0	28
relief	126	28	28	28.0	28
remorse	474	28	28	28.0	28
sadness	1129	28	28	28.0	28
surprise	932	28	28	28.0	28

Source fields by label class (train)

Scalar source fields with 2–60 values. 'Purity' = accuracy of predicting the label class from this field alone (per-value majority), against the overall majority. The model does not see source, but a field that predicts the label marks a confound: rows of one origin carry one label, so any style difference of that origin becomes a shortcut.

Overall majority: 31.3%.

source field	values	purity	top values → classes
target_kind	2	31.3%	hard: neutral 35.6%, admiration 7.4%; soft: neutral 9.9%, admiration 9.6%
instructions	22	31.3%	canonical: neutral 31.2%, admiration 7.8%; none: neutral 30.9%, admiration 7.4%; variant-18: neutral 31.7%, admiration 7.5%; variant-14: neutral 31.6%, admiration 9.5%; variant-12: neutral 32.5%, admiration 7.8%; variant-19: neutral 29.4%, admiration 8.3%
split	2	31.3%	train: neutral 31.3%, admiration 7.8%; validation: neutral 31.6%, admiration 7.5%
copies	9	31.3%	1: neutral 31.3%, admiration 7.8%; 2: neutral 28.8%, gratitude 15.8%; 3: neutral 27.5%, gratitude 15.0%; 4: gratitude 27.3%, neutral 27.3%; 5: neutral 44.4%, surprise 11.1%; 6: neutral 50.0%, joy 25.0%

Formatting by label class (main text, share of rows)

feature	lowest classes	highest classes
ends with ?	pride 0%, relief 1%, joy 1%	curiosity 51%, confusion 32%, neutral 5%	gap
ends with .	curiosity 25%, amusement 34%, excitement 35%	disapproval 62%, realization 58%, approval 57%	gap
ends with !	grief 1%, remorse 3%, confusion 3%	excitement 29%, gratitude 23%, joy 15%	gap
no end punctuation	curiosity 18%, confusion 25%, caring 27%	amusement 54%, grief 43%, sadness 41%	gap
starts lowercase	relief 4%, desire 4%, realization 4%	grief 9%, amusement 9%, nervousness 8%
all lowercase	pride 1%, relief 2%, desire 2%	embarrassment 7%, nervousness 6%, grief 6%
has a digit	anger 4%, caring 5%, gratitude 5%	disappointment 12%, surprise 12%, realization 12%
has newline	admiration 0%, amusement 0%, anger 0%	surprise 0%, sadness 0%, remorse 0%
has quotes	relief 1%, grief 1%, optimism 2%	annoyance 6%, amusement 5%, embarrassment 5%
has markup (HTML/markdown)	caring 0%, desire 0%, disgust 0%	anger 1%, relief 1%, joy 1%
has URL	admiration 0%, amusement 0%, anger 0%	surprise 0%, sadness 0%, remorse 0%
non-ASCII	neutral 12%, anger 12%, desire 13%	relief 21%, disapproval 21%, remorse 20%
non-Latin script	admiration 0%, anger 0%, annoyance 0%	embarrassment 0%, disapproval 0%, surprise 0%
emoji	pride 0%, relief 0%, anger 1%	grief 4%, love 4%, sadness 4%
ALL-CAPS word (4+)	embarrassment 8%, nervousness 9%, gratitude 10%	love 23%, anger 23%, neutral 21%
contains ' - ' or —	grief 0%, pride 0%, sadness 0%	nervousness 2%, gratitude 1%, embarrassment 1%

Same, by row kind

feature	emotion
ends with ?	6%
ends with .	48%
ends with !	9%
no end punctuation	34%
starts lowercase	6%
all lowercase	4%
has a digit	9%
has newline	0%
has quotes	4%
has markup (HTML/markdown)	0%
has URL	0%
non-ASCII	15%
non-Latin script	0%
emoji	2%
ALL-CAPS word (4+)	17%
contains ' - ' or —	1%

Over-represented words and phrases per label class (main text)

Log-odds ratio with an informative Dirichlet prior (Monroe et al. 2008), each class against all the others; z-score, then the share of rows in the class and in the other classes that contain the phrase. Counted once per row.

Words, admiration: great z=42 14.1% vs 0.8%; good z=32 14.1% vs 2.9%; awesome z=29 6.7% vs 0.2%; amazing z=28 6.0% vs 0.3%; nice z=24 5.4% vs 0.6%; beautiful z=22 4.1% vs 0.1%; best z=20 5.3% vs 1.0%; pretty z=20 5.4% vs 1.1%; cute z=18 2.6% vs 0.2%; appreciate z=16 2.2% vs 0.2%; looks z=14 3.7% vs 1.0%; cool z=13 2.6% vs 0.6%; fantastic z=12 1.2% vs 0.1%; job z=12 2.0% vs 0.4%; wow z=11 2.9% vs 1.0%; is z=11 23.0% vs 17.3%; wonderful z=11 1.0% vs 0.1%; incredible z=10 0.8% vs 0.0%; excellent z=10 0.8% vs 0.0%; interesting z=10 1.8% vs 0.5%

Words, amusement: lol z=64 42.6% vs 0.9%; haha z=33 11.0% vs 0.3%; funny z=30 9.1% vs 0.3%; lmao z=23 5.6% vs 0.3%; fun z=22 6.2% vs 0.6%; hilarious z=18 3.2% vs 0.1%; joke z=17 3.6% vs 0.3%; laugh z=17 2.9% vs 0.2%; hahaha z=16 2.8% vs 0.1%; laughed z=14 2.1% vs 0.0%; laughing z=12 1.6% vs 0.1%; ha z=8 1.0% vs 0.1%; loud z=7 0.8% vs 0.1%; kidding z=7 0.6% vs 0.0%; jokes z=7 0.7% vs 0.1%; joking z=7 0.5% vs 0.0%; hahahaha z=7 0.5% vs 0.0%; funnier z=7 0.5% vs 0.0%; entertaining z=6 0.5% vs 0.0%; actually z=6 3.2% vs 1.5%

Words, anger: fuck z=36 15.0% vs 0.5%; hate z=27 10.2% vs 0.6%; fucking z=26 8.8% vs 0.5%; angry z=15 2.8% vs 0.1%; stupid z=14 3.8% vs 0.5%; dare z=14 2.2% vs 0.1%; shut z=12 2.1% vs 0.1%; hell z=12 3.2% vs 0.5%; shit z=9 3.2% vs 0.7%; asshole z=9 1.0% vs 0.0%; fucked z=9 1.1% vs 0.1%; wtf z=9 1.1% vs 0.1%; bitch z=9 1.1% vs 0.1%; idiot z=9 1.4% vs 0.2%; bastard z=9 0.9% vs 0.0%; bullshit z=8 1.0% vs 0.1%; stop z=8 2.7% vs 0.7%; kill z=8 1.5% vs 0.3%; suck z=7 1.0% vs 0.1%; off z=7 3.2% vs 1.3%

Words, annoyance: stupid z=16 3.8% vs 0.4%; annoying z=13 1.9% vs 0.0%; fucking z=13 3.7% vs 0.6%; shit z=12 3.6% vs 0.7%; damn z=11 3.7% vs 0.9%; dumb z=10 1.7% vs 0.2%; idiot z=10 1.4% vs 0.1%; fuck z=9 3.1% vs 0.8%; sucks z=9 1.7% vs 0.3%; annoyed z=8 0.8% vs 0.0%; idiots z=8 0.9% vs 0.1%; weird z=8 2.2% vs 0.5%; stop z=8 2.5% vs 0.7%; pissed z=8 0.9% vs 0.1%; hate z=7 2.6% vs 0.8%; hell z=7 1.8% vs 0.5%; ridiculous z=7 0.9% vs 0.1%; asshole z=7 0.6% vs 0.0%; fool z=6 0.5% vs 0.0%; frustrating z=6 0.5% vs 0.0%

Words, approval: agree z=26 6.4% vs 0.2%; yes z=16 5.1% vs 0.9%; yeah z=13 5.7% vs 1.7%; right z=12 5.9% vs 1.9%; agreed z=12 1.4% vs 0.1%; true z=11 2.7% vs 0.6%; correct z=9 1.2% vs 0.1%; ok z=9 2.4% vs 0.6%; sure z=9 3.8% vs 1.4%; exactly z=8 1.8% vs 0.4%; yep z=7 0.9% vs 0.2%; definitely z=7 1.8% vs 0.6%; fine z=7 1.1% vs 0.3%; fair z=6 0.9% vs 0.2%; but z=6 12.0% vs 8.2%; free z=5 1.1% vs 0.4%; prefer z=5 0.5% vs 0.1%; it's z=5 6.5% vs 4.2%; with z=5 9.9% vs 6.9%; absolutely z=5 1.1% vs 0.4%

Words, caring: worry z=17 4.8% vs 0.2%; you z=16 47.3% vs 19.1%; yourself z=16 6.2% vs 0.5%; your z=14 19.1% vs 5.5%; stay z=14 4.5% vs 0.3%; safe z=13 3.2% vs 0.2%; luck z=13 4.8% vs 0.5%; help z=13 5.8% vs 0.8%; careful z=11 2.1% vs 0.0%; concerned z=11 1.8% vs 0.0%; care z=10 3.7% vs 0.5%; bless z=10 1.8% vs 0.1%; better z=9 5.9% vs 1.5%; get z=9 11.0% vs 4.0%; take z=9 4.8% vs 1.1%; need z=8 5.2% vs 1.3%; keep z=8 4.0% vs 0.9%; strong z=8 1.9% vs 0.2%; please z=7 3.2% vs 0.7%; praying z=7 0.9% vs 0.0%

Words, confusion: sure z=19 10.2% vs 1.3%; confused z=19 5.9% vs 0.0%; why z=17 11.5% vs 2.0%; or z=15 12.5% vs 3.0%; what z=14 17.6% vs 5.6%; understand z=14 4.6% vs 0.5%; don't z=12 10.9% vs 3.1%; know z=12 10.6% vs 3.0%; idea z=10 3.5% vs 0.5%; idk z=10 2.1% vs 0.2%; not z=10 17.7% vs 7.9%; how z=8 9.4% vs 3.7%; confusing z=8 0.9% vs 0.0%; maybe z=8 3.9% vs 1.0%; confusion z=7 0.8% vs 0.0%; don z=7 4.4% vs 1.4%; doubt z=7 1.2% vs 0.1%; clue z=6 0.7% vs 0.0%; do z=6 8.6% vs 4.2%; i z=6 45.8% vs 31.7%

Words, curiosity: what z=25 20.9% vs 5.2%; curious z=21 6.5% vs 0.1%; why z=17 9.0% vs 2.0%; how z=17 12.1% vs 3.4%; did z=16 7.8% vs 1.8%; do z=14 11.4% vs 4.0%; you z=13 33.0% vs 19.1%; does z=12 4.3% vs 1.0%; what's z=11 2.0% vs 0.2%; where z=11 4.0% vs 1.0%; are z=10 13.2% vs 6.8%; curiosity z=9 1.0% vs 0.0%; interesting z=8 2.2% vs 0.5%; anyone z=7 2.4% vs 0.7%; wonder z=7 1.4% vs 0.3%; explain z=7 0.9% vs 0.1%; wondering z=6 0.8% vs 0.1%; question z=6 1.2% vs 0.3%; any z=6 3.2% vs 1.4%; or z=6 5.7% vs 3.1%

Words, desire: wish z=44 39.7% vs 0.5%; want z=17 14.8% vs 1.7%; i z=15 74.5% vs 31.5%; could z=14 11.6% vs 1.5%; wanted z=12 4.8% vs 0.3%; need z=9 7.2% vs 1.3%; wanna z=7 2.2% vs 0.2%; i'd z=6 3.7% vs 0.7%; would z=6 10.3% vs 3.8%; hope z=6 5.9% vs 1.8%; more z=6 8.1% vs 3.0%; dream z=5 1.3% vs 0.1%; hadn't z=5 0.7% vs 0.0%; pray z=5 0.9% vs 0.1%; had z=5 5.9% vs 2.2%; to z=5 37.8% vs 24.7%; desire z=5 0.6% vs 0.0%; please z=4 2.6% vs 0.7%; expecting z=4 0.9% vs 0.1%; blunder z=4 0.4% vs 0.0%

Words, disappointment: disappointed z=15 3.3% vs 0.1%; bad z=14 9.0% vs 1.6%; disappointing z=9 2.4% vs 0.0%; upset z=9 1.6% vs 0.1%; depressing z=9 1.1% vs 0.0%; lost z=9 2.5% vs 0.3%; miss z=9 2.3% vs 0.3%; unfortunately z=8 1.8% vs 0.2%; disappointment z=8 1.2% vs 0.0%; missed z=8 1.7% vs 0.2%; tried z=7 1.8% vs 0.3%; hurt z=7 1.7% vs 0.2%; game z=7 4.2% vs 1.3%; upsetting z=6 0.7% vs 0.0%; lose z=6 1.5% vs 0.3%; poor z=6 1.9% vs 0.4%; disaster z=6 0.6% vs 0.0%; sucks z=5 1.4% vs 0.3%; worse z=5 1.6% vs 0.4%; boring z=5 0.8% vs 0.1%

Words, disapproval: not z=22 24.8% vs 7.5%; no z=18 14.4% vs 3.9%; don't z=16 11.1% vs 2.9%; t z=15 12.9% vs 4.0%; don z=13 5.9% vs 1.3%; disagree z=12 1.6% vs 0.1%; nope z=10 1.3% vs 0.1%; doesn't z=9 3.3% vs 0.9%; wrong z=9 3.1% vs 0.8%; can't z=9 3.9% vs 1.2%; nah z=8 1.5% vs 0.2%; isn't z=8 2.6% vs 0.7%; unpopular z=8 0.7% vs 0.0%; think z=8 7.0% vs 3.1%; dont z=8 2.1% vs 0.5%; doesn z=7 1.5% vs 0.3%; bad z=6 4.0% vs 1.7%; refuse z=6 0.5% vs 0.0%; illegal z=6 0.7% vs 0.1%; opinion z=6 1.1% vs 0.2%

Words, disgust: awful z=23 9.6% vs 0.2%; disgusting z=22 13.5% vs 0.0%; worst z=22 9.2% vs 0.3%; weird z=17 7.6% vs 0.5%; worse z=15 5.8% vs 0.3%; nasty z=11 2.2% vs 0.0%; ugly z=10 2.5% vs 0.1%; gross z=10 2.1% vs 0.1%; creepy z=10 2.2% vs 0.1%; terrible z=8 2.7% vs 0.4%; ugh z=7 1.6% vs 0.1%; is z=7 28.4% vs 17.6%; horrible z=7 1.9% vs 0.3%; hideous z=6 0.7% vs 0.0%; dirty z=6 0.9% vs 0.1%; disgusted z=6 1.2% vs 0.0%; disgust z=6 0.6% vs 0.0%; ever z=6 3.4% vs 1.0%; filthy z=6 0.6% vs 0.0%; bad z=5 4.7% vs 1.7%

Words, embarrassment: shame z=18 11.5% vs 0.2%; awkward z=18 10.4% vs 0.1%; embarrassing z=16 11.5% vs 0.0%; ashamed z=12 5.0% vs 0.0%; embarrassed z=12 4.6% vs 0.0%; embarrassment z=10 5.0% vs 0.0%; oops z=8 1.9% vs 0.0%; weird z=7 5.4% vs 0.6%; uncomfortable z=7 2.3% vs 0.1%; cringy z=7 1.5% vs 0.0%; shamed z=5 0.8% vs 0.0%; temporarily z=5 0.8% vs 0.0%; millionaires z=5 0.8% vs 0.0%; forgot z=5 2.3% vs 0.3%; translate z=4 0.8% vs 0.0%; shaming z=4 0.8% vs 0.0%; cringey z=4 0.8% vs 0.0%; feel z=4 5.8% vs 1.6%; socially z=4 0.8% vs 0.0%; abused z=4 0.8% vs 0.0%

Words, excitement: excited z=25 11.1% vs 0.1%; wait z=16 6.6% vs 0.6%; wow z=15 7.8% vs 1.0%; interesting z=14 5.8% vs 0.5%; happy z=13 7.4% vs 1.1%; cake z=11 2.6% vs 0.1%; birthday z=11 2.3% vs 0.1%; exciting z=10 1.6% vs 0.0%; yay z=9 1.4% vs 0.0%; new z=9 4.8% vs 1.0%; omg z=7 2.5% vs 0.4%; interested z=7 1.6% vs 0.2%; excitement z=7 0.9% vs 0.0%; day z=7 4.0% vs 1.2%; year z=6 3.8% vs 1.1%; crazy z=6 2.0% vs 0.4%; amazing z=6 2.7% vs 0.7%; stoked z=6 0.6% vs 0.0%; cheers z=6 1.3% vs 0.2%; cakeday z=6 0.7% vs 0.0%

Words, fear: afraid z=23 10.7% vs 0.1%; scared z=22 11.1% vs 0.1%; terrible z=21 10.0% vs 0.3%; horrible z=19 7.5% vs 0.2%; scary z=18 6.8% vs 0.0%; terrifying z=15 5.3% vs 0.0%; fear z=15 4.4% vs 0.1%; dangerous z=11 2.6% vs 0.1%; worried z=11 3.0% vs 0.1%; creepy z=10 2.8% vs 0.1%; cringe z=10 2.8% vs 0.1%; horrifying z=10 1.9% vs 0.0%; horror z=9 1.8% vs 0.0%; nightmare z=8 1.6% vs 0.1%; scares z=8 2.3% vs 0.0%; me z=8 15.6% vs 6.3%; horrific z=7 1.1% vs 0.0%; terrified z=6 1.6% vs 0.0%; horribly z=6 0.9% vs 0.0%; frightening z=6 0.7% vs 0.0%

Words, gratitude: thanks z=63 47.3% vs 0.5%; thank z=56 38.2% vs 0.4%; for z=37 39.7% vs 11.1%; you z=29 45.1% vs 18.3%; sharing z=17 2.9% vs 0.1%; advice z=16 2.5% vs 0.1%; appreciate z=15 2.7% vs 0.2%; i'll z=15 3.5% vs 0.5%; much z=13 6.6% vs 2.1%; welcome z=13 2.0% vs 0.2%; info z=12 1.5% vs 0.1%; congrats z=11 1.5% vs 0.2%; glad z=11 3.4% vs 1.0%; helpful z=10 1.1% vs 0.1%; very z=9 4.4% vs 1.7%; ll z=9 2.2% vs 0.5%; response z=9 1.2% vs 0.2%; luck z=9 2.0% vs 0.5%; tip z=9 0.8% vs 0.1%; reply z=8 1.1% vs 0.1%

Words, grief: died z=15 22.4% vs 0.1%; rip z=10 10.4% vs 0.1%; loss z=9 10.4% vs 0.2%; dead z=8 10.4% vs 0.2%; death z=7 9.0% vs 0.2%; heart z=6 7.5% vs 0.2%; deaths z=6 3.0% vs 0.0%; peace z=6 4.5% vs 0.1%; condolences z=6 3.0% vs 0.0%; sooner z=5 3.0% vs 0.0%; sorry z=5 14.9% vs 1.7%; disease z=5 3.0% vs 0.0%; passed z=4 3.0% vs 0.1%; lost z=4 6.0% vs 0.4%; lands z=4 1.5% vs 0.0%; depression z=4 3.0% vs 0.1%; 2013 z=4 1.5% vs 0.0%; grandfather z=4 1.5% vs 0.0%; windshield z=4 1.5% vs 0.0%; chills z=4 1.5% vs 0.0%

Words, joy: happy z=40 21.2% vs 0.7%; glad z=35 17.3% vs 0.6%; enjoy z=26 8.7% vs 0.2%; fun z=18 7.0% vs 0.7%; enjoyed z=14 2.7% vs 0.0%; cheers z=12 2.2% vs 0.1%; enjoying z=11 1.6% vs 0.0%; cake z=11 1.9% vs 0.1%; day z=10 4.9% vs 1.1%; joy z=10 1.3% vs 0.0%; i'm z=10 10.4% vs 4.0%; happiness z=8 1.0% vs 0.1%; cakeday z=7 0.8% vs 0.0%; so z=7 13.6% vs 7.6%; m z=7 5.4% vs 2.1%; smile z=7 1.1% vs 0.1%; happily z=7 0.6% vs 0.0%; made z=6 3.3% vs 1.1%; birthday z=6 0.9% vs 0.1%; gladly z=6 0.5% vs 0.0%

Words, love: love z=81 69.0% vs 1.6%; i z=28 66.3% vs 30.6%; loved z=22 5.2% vs 0.2%; favorite z=17 4.0% vs 0.4%; loves z=17 3.0% vs 0.1%; like z=12 14.5% vs 7.0%; loving z=12 1.4% vs 0.1%; my z=10 14.7% vs 7.8%; i'd z=9 2.7% vs 0.7%; name z=7 19.9% vs 14.0%; lovely z=7 0.8% vs 0.1%; favourite z=6 0.7% vs 0.1%; liked z=6 0.9% vs 0.2%; d z=6 2.0% vs 0.7%; this z=6 18.4% vs 13.9%; much z=6 4.2% vs 2.2%; song z=5 0.8% vs 0.2%; sweet z=5 1.0% vs 0.3%; it z=5 21.7% vs 17.2%; how z=5 6.0% vs 3.7%

Words, nervousness: worried z=15 13.8% vs 0.1%; nervous z=14 11.5% vs 0.0%; anxiety z=12 9.2% vs 0.1%; anxious z=11 6.9% vs 0.0%; worrying z=9 4.6% vs 0.0%; worry z=7 5.4% vs 0.2%; career z=6 3.1% vs 0.1%; stressed z=5 1.5% vs 0.0%; scary z=5 3.1% vs 0.1%; paranoid z=5 1.5% vs 0.0%; breathing z=5 1.5% vs 0.0%; nightmares z=5 1.5% vs 0.0%; about z=5 16.2% vs 4.3%; that'll z=4 1.5% vs 0.0%; panic z=4 1.5% vs 0.0%; me z=4 20.0% vs 6.4%; edgy z=4 1.5% vs 0.0%; pick z=4 3.1% vs 0.2%; i'm z=4 14.6% vs 4.2%; standing z=4 1.5% vs 0.0%

Words, neutral: they z=15 8.2% vs 5.0%; name z=15 17.3% vs 12.8%; he z=9 6.9% vs 5.0%; in z=8 14.1% vs 12.1%; on z=7 8.8% vs 7.4%; only z=7 2.8% vs 1.9%; by z=7 2.6% vs 1.8%; the z=7 33.8% vs 32.3%; she z=6 3.6% vs 2.7%; his z=6 3.3% vs 2.5%; or z=6 3.8% vs 3.0%; their z=6 2.5% vs 1.9%; left z=6 0.8% vs 0.4%; 1 z=5 0.9% vs 0.5%; r z=5 0.9% vs 0.5%; from z=5 3.9% vs 3.2%; who z=5 2.8% vs 2.2%; then z=5 2.4% vs 1.8%; go z=5 2.5% vs 1.9%; there's z=5 0.7% vs 0.4%

Words, optimism: hope z=51 36.0% vs 0.8%; luck z=21 7.3% vs 0.4%; hopefully z=21 6.8% vs 0.1%; hoping z=19 5.1% vs 0.1%; will z=13 10.2% vs 2.5%; good z=10 10.5% vs 3.6%; soon z=10 2.2% vs 0.2%; better z=9 5.6% vs 1.5%; wish z=7 3.4% vs 0.8%; can z=7 10.1% vs 4.2%; next z=7 2.7% vs 0.6%; i z=7 50.7% vs 31.5%; best z=7 4.3% vs 1.3%; win z=7 1.9% vs 0.3%; future z=7 1.5% vs 0.2%; goes z=7 1.6% vs 0.3%; probably z=6 3.3% vs 1.0%; get z=6 8.8% vs 4.0%; hopeful z=6 0.5% vs 0.0%; optimistic z=6 0.5% vs 0.0%

Words, pride: proud z=23 38.3% vs 0.1%; pride z=9 6.2% vs 0.0%; jersey z=5 3.7% vs 0.1%; winner z=5 2.5% vs 0.0%; myself z=5 6.2% vs 0.5%; am z=4 9.9% vs 1.3%; butterfly z=4 1.2% vs 0.0%; cubs z=4 1.2% vs 0.0%; european z=4 1.2% vs 0.0%; honored z=4 1.2% vs 0.0%; leaking z=4 1.2% vs 0.0%; integrity z=4 1.2% vs 0.0%; congrats z=4 3.7% vs 0.2%; corn z=4 1.2% vs 0.0%; playthrough z=4 1.2% vs 0.0%; accomplishment z=4 1.2% vs 0.0%; senator z=4 1.2% vs 0.0%; athletic z=4 1.2% vs 0.0%; 2006 z=4 1.2% vs 0.0%; glory z=4 1.2% vs 0.0%

Words, realization: realize z=18 5.8% vs 0.2%; realized z=16 4.2% vs 0.0%; thought z=10 6.1% vs 1.2%; forgot z=10 2.7% vs 0.2%; was z=9 19.3% vs 8.2%; reminds z=8 1.7% vs 0.1%; noticed z=8 1.8% vs 0.2%; until z=8 3.1% vs 0.6%; realizing z=7 0.9% vs 0.0%; realised z=7 0.9% vs 0.0%; figured z=7 1.0% vs 0.1%; ago z=7 2.7% vs 0.5%; realise z=7 0.8% vs 0.0%; didn't z=6 4.3% vs 1.2%; realization z=6 0.6% vs 0.0%; i z=6 48.7% vs 31.7%; aware z=5 0.9% vs 0.1%; wrong z=5 2.9% vs 0.8%; that z=5 28.6% vs 17.9%; ve z=5 2.7% vs 0.8%

Words, relief: glad z=14 22.2% vs 1.0%; relief z=9 4.8% vs 0.0%; god z=8 7.1% vs 0.3%; relieved z=8 3.2% vs 0.0%; least z=8 10.3% vs 0.8%; finally z=6 4.8% vs 0.2%; relax z=6 2.4% vs 0.0%; whew z=5 1.6% vs 0.0%; feel z=5 9.5% vs 1.6%; cool z=5 6.3% vs 0.8%; solved z=5 1.6% vs 0.0%; thank z=5 11.1% vs 2.3%; thankfully z=5 1.6% vs 0.0%; mac z=5 1.6% vs 0.0%; safe z=4 3.2% vs 0.2%; could've z=4 1.6% vs 0.0%; goodness z=4 1.6% vs 0.0%; m z=4 9.5% vs 2.2%; at z=4 15.1% vs 4.7%; oof z=4 1.6% vs 0.1%

Words, remorse: sorry z=59 75.9% vs 0.9%; regret z=16 5.5% vs 0.0%; apologies z=12 3.4% vs 0.0%; m z=11 11.4% vs 2.1%; i'm z=10 15.2% vs 4.1%; apologize z=9 1.9% vs 0.0%; guilty z=8 1.7% vs 0.0%; meant z=8 3.0% vs 0.2%; guilt z=7 1.3% vs 0.0%; i z=7 54.4% vs 31.8%; loss z=7 2.1% vs 0.2%; im z=6 3.4% vs 0.5%; am z=6 5.1% vs 1.3%; apology z=5 0.6% vs 0.0%; misunderstanding z=5 0.6% vs 0.0%; happened z=5 2.5% vs 0.5%; sincerely z=5 0.6% vs 0.0%; didn z=5 2.7% vs 0.6%; misread z=5 0.6% vs 0.0%; should've z=5 0.6% vs 0.0%

Words, sadness: sad z=35 17.3% vs 0.2%; sorry z=23 12.8% vs 1.4%; sadly z=18 5.0% vs 0.0%; miss z=14 3.5% vs 0.3%; hurts z=13 2.4% vs 0.1%; painful z=13 2.3% vs 0.0%; poor z=13 3.5% vs 0.3%; feel z=12 6.9% vs 1.5%; crying z=12 2.2% vs 0.1%; pain z=11 2.3% vs 0.2%; bad z=11 6.8% vs 1.6%; cry z=11 1.9% vs 0.1%; loss z=9 1.8% vs 0.1%; lonely z=8 1.0% vs 0.0%; my z=8 15.1% vs 7.9%; hurt z=8 1.7% vs 0.2%; i'm z=8 9.2% vs 4.1%; lost z=7 1.9% vs 0.3%; hard z=7 3.2% vs 0.8%; sick z=7 1.3% vs 0.2%

Words, surprise: wow z=31 17.0% vs 0.8%; surprised z=30 14.2% vs 0.1%; wonder z=22 7.6% vs 0.2%; omg z=18 6.1% vs 0.3%; wondering z=14 3.2% vs 0.1%; shocked z=14 3.2% vs 0.0%; oh z=14 9.4% vs 1.9%; surprise z=14 2.8% vs 0.1%; believe z=13 5.2% vs 0.7%; surprising z=10 1.4% vs 0.0%; unexpected z=8 1.1% vs 0.0%; wondered z=8 1.1% vs 0.0%; strange z=8 1.3% vs 0.1%; was z=8 16.1% vs 8.2%; amazed z=8 0.9% vs 0.0%; unbelievable z=8 0.9% vs 0.0%; god z=7 2.0% vs 0.3%; shocking z=7 0.8% vs 0.0%; how z=7 8.5% vs 3.7%; holy z=7 1.5% vs 0.2%

2–4-word phrases, admiration: a great z=22 3.9% vs 0.2%; the best z=19 3.8% vs 0.5%; a good z=15 2.8% vs 0.5%; i appreciate z=13 1.4% vs 0.1%; this is z=13 6.0% vs 2.5%; is amazing z=12 1.2% vs 0.0%; is awesome z=12 1.3% vs 0.0%; is great z=12 1.1% vs 0.1%; what a z=11 2.1% vs 0.5%; is the best z=11 0.9% vs 0.1%; good job z=10 0.9% vs 0.0%; such a z=10 1.6% vs 0.4%; so good z=10 0.8% vs 0.1%; a beautiful z=10 0.8% vs 0.0%; pretty good z=9 0.7% vs 0.0%; is a great z=9 0.7% vs 0.0%; so cute z=9 0.6% vs 0.0%; an amazing z=9 0.6% vs 0.0%; a pretty z=8 0.7% vs 0.1%; a nice z=8 0.8% vs 0.1%

2–4-word phrases, amusement: lol i z=16 2.7% vs 0.1%; haha i z=12 1.4% vs 0.0%; me laugh z=11 1.3% vs 0.0%; a joke z=10 1.3% vs 0.1%; made me laugh z=10 1.0% vs 0.0%; i laughed z=10 1.2% vs 0.0%; it lol z=9 0.9% vs 0.0%; fun to z=8 0.7% vs 0.0%; so hard z=8 1.0% vs 0.1%; is hilarious z=8 0.8% vs 0.0%; name lol z=8 0.7% vs 0.0%; lol you z=8 0.7% vs 0.0%; made me z=8 1.4% vs 0.3%; out loud z=8 0.6% vs 0.0%; it's funny z=8 0.6% vs 0.0%; funny how z=8 0.6% vs 0.0%; at this z=7 1.2% vs 0.2%; lol the z=7 0.6% vs 0.0%; lol this z=7 0.7% vs 0.0%; the joke z=7 0.6% vs 0.1%

2–4-word phrases, anger: i hate z=22 6.0% vs 0.3%; the fuck z=19 4.2% vs 0.1%; how dare z=13 2.0% vs 0.0%; what the z=12 2.6% vs 0.3%; what the fuck z=11 1.4% vs 0.0%; the hell z=11 1.5% vs 0.1%; dare you z=11 1.3% vs 0.0%; fuck you z=10 1.4% vs 0.0%; fuck is z=10 1.2% vs 0.0%; fuck off z=10 1.3% vs 0.0%; shut up z=10 1.1% vs 0.0%; how dare you z=10 1.3% vs 0.0%; the fuck is z=10 1.1% vs 0.0%; fuck the z=9 1.2% vs 0.0%; the fucking z=9 0.9% vs 0.0%; a fucking z=9 1.0% vs 0.1%; hate the z=8 0.9% vs 0.0%; you fucking z=8 0.7% vs 0.0%; hate that z=8 0.7% vs 0.0%; what the hell z=8 0.7% vs 0.0%

2–4-word phrases, annoyance: an idiot z=9 0.9% vs 0.0%; a fucking z=7 0.7% vs 0.1%; that sucks z=7 0.5% vs 0.0%; the hell z=6 0.8% vs 0.1%; an asshole z=6 0.4% vs 0.0%; a weird z=6 0.6% vs 0.1%; i hate z=6 1.3% vs 0.4%; can't even z=6 0.4% vs 0.0%; stupid to z=6 0.3% vs 0.0%; what the hell z=5 0.4% vs 0.0%; shut up z=5 0.4% vs 0.1%; is stupid z=5 0.3% vs 0.0%; fuck that z=5 0.3% vs 0.0%; a fuck z=5 0.3% vs 0.0%; waste of z=5 0.3% vs 0.0%; makes no z=5 0.3% vs 0.0%; as hell z=5 0.5% vs 0.1%; tired of z=5 0.3% vs 0.0%; a god z=5 0.3% vs 0.0%; a stupid z=5 0.3% vs 0.0%

2–4-word phrases, approval: i agree z=18 3.8% vs 0.1%; agree with z=14 1.8% vs 0.1%; i agree with z=10 1.0% vs 0.0%; yeah i z=8 1.5% vs 0.3%; agree with you z=8 0.6% vs 0.0%; you're right z=8 0.7% vs 0.1%; pretty sure z=7 0.7% vs 0.1%; you re right z=7 0.6% vs 0.0%; right i z=7 0.7% vs 0.1%; re right z=7 0.6% vs 0.0%; agree that z=7 0.5% vs 0.0%; with you z=7 1.1% vs 0.2%; i think z=6 3.3% vs 1.5%; yes i z=6 0.8% vs 0.2%; agree but z=6 0.4% vs 0.0%; for sure z=6 0.7% vs 0.1%; agree i z=6 0.5% vs 0.0%; yes that z=6 0.4% vs 0.0%; i believe z=6 0.6% vs 0.1%; i completely z=6 0.4% vs 0.0%

2–4-word phrases, caring: don't worry z=12 2.5% vs 0.0%; for you z=12 4.7% vs 0.6%; good luck z=11 3.6% vs 0.3%; you need z=11 2.9% vs 0.2%; take care z=10 1.6% vs 0.0%; be careful z=10 1.5% vs 0.0%; stay strong z=9 1.5% vs 0.0%; stay safe z=9 1.8% vs 0.0%; name bless z=8 1.2% vs 0.0%; care of z=8 1.3% vs 0.1%; keep your z=8 1.1% vs 0.0%; you have z=8 3.8% vs 0.8%; if you z=8 4.9% vs 1.2%; t worry z=8 1.0% vs 0.0%; don t worry z=8 1.0% vs 0.0%; don't be z=7 1.0% vs 0.0%; take care of z=7 0.9% vs 0.0%; you feel z=7 1.4% vs 0.1%; get some z=7 1.1% vs 0.1%; no worries z=7 0.9% vs 0.0%

2–4-word phrases, confusion: not sure z=23 7.8% vs 0.2%; i don't z=15 7.7% vs 1.0%; don't know z=14 4.1% vs 0.3%; i don't know z=13 3.1% vs 0.2%; no idea z=13 2.9% vs 0.1%; sure what z=12 2.3% vs 0.0%; i'm not sure z=11 1.9% vs 0.0%; don t know z=11 2.1% vs 0.1%; not sure what z=11 1.7% vs 0.0%; have no idea z=11 1.8% vs 0.1%; i don t know z=10 1.7% vs 0.1%; i have no idea z=10 1.6% vs 0.0%; i have no z=10 1.8% vs 0.1%; t know z=10 2.4% vs 0.2%; m not sure z=10 1.4% vs 0.0%; i m not sure z=10 1.4% vs 0.0%; sure if z=10 1.4% vs 0.0%; i don z=9 3.5% vs 0.5%; i don t z=9 3.5% vs 0.5%; know what z=9 2.6% vs 0.3%

2–4-word phrases, curiosity: do you z=22 6.5% vs 0.5%; are you z=18 4.9% vs 0.5%; did you z=16 3.2% vs 0.2%; is it z=14 2.7% vs 0.2%; is this z=13 2.7% vs 0.3%; can you z=11 1.7% vs 0.1%; what are z=11 1.4% vs 0.1%; is that z=10 2.8% vs 0.5%; do you have z=10 1.2% vs 0.1%; just curious z=10 1.4% vs 0.0%; have you z=10 1.3% vs 0.1%; how do z=9 1.1% vs 0.1%; you think z=9 1.8% vs 0.3%; how did z=9 1.0% vs 0.0%; what do z=9 1.1% vs 0.1%; why is z=9 1.1% vs 0.1%; what are you z=9 0.9% vs 0.0%; how is z=9 0.9% vs 0.0%; is there z=9 1.2% vs 0.1%; do you think z=9 1.0% vs 0.1%

2–4-word phrases, desire: i wish z=36 29.9% vs 0.2%; wish i z=24 13.1% vs 0.1%; i wish i z=22 10.9% vs 0.1%; wish i could z=17 6.6% vs 0.1%; i want z=17 8.5% vs 0.4%; i could z=16 7.7% vs 0.3%; i wish i could z=16 5.5% vs 0.1%; i want to z=12 4.2% vs 0.2%; want to z=10 6.6% vs 0.8%; wish we z=10 2.6% vs 0.0%; i just want z=9 2.2% vs 0.1%; wish i had z=9 2.0% vs 0.0%; i wanna z=9 2.2% vs 0.1%; wish the z=9 2.4% vs 0.0%; i need z=9 3.0% vs 0.2%; just want z=9 2.2% vs 0.1%; wish they z=8 1.7% vs 0.0%; wish it z=8 1.7% vs 0.0%; wish name z=8 1.3% vs 0.0%; i wish name z=8 1.3% vs 0.0%

2–4-word phrases, disappointment: my bad z=8 1.4% vs 0.1%; i miss z=8 1.5% vs 0.1%; i tried z=7 1.0% vs 0.1%; too bad z=7 1.1% vs 0.1%; disappointed in z=7 0.8% vs 0.0%; bad for z=6 1.0% vs 0.1%; missed it z=6 0.6% vs 0.0%; i missed z=6 0.8% vs 0.1%; so bad z=6 0.9% vs 0.1%; i tried to z=6 0.6% vs 0.0%; a disaster z=6 0.5% vs 0.0%; i lost z=5 0.6% vs 0.0%; oh no z=5 0.8% vs 0.1%; it didn z=5 0.5% vs 0.0%; it didn t z=5 0.5% vs 0.0%; tried to z=5 0.8% vs 0.1%; but no z=5 0.6% vs 0.0%; so disappointed z=5 0.4% vs 0.0%; filled with z=5 0.5% vs 0.0%; this game z=5 0.9% vs 0.2%

2–4-word phrases, disapproval: i don't z=14 5.6% vs 1.0%; don t z=13 5.9% vs 1.3%; don't think z=12 2.3% vs 0.2%; i don z=12 3.2% vs 0.5%; i don t z=12 3.2% vs 0.5%; that's not z=12 1.8% vs 0.1%; is not z=12 2.7% vs 0.4%; i don't think z=11 2.0% vs 0.2%; not a z=10 2.5% vs 0.4%; it's not z=10 2.3% vs 0.4%; i dont z=9 1.5% vs 0.2%; don t think z=8 1.1% vs 0.1%; no it z=8 0.8% vs 0.0%; t think z=8 1.2% vs 0.1%; don t want z=8 0.7% vs 0.0%; disagree with z=8 0.7% vs 0.0%; don't like z=8 0.9% vs 0.1%; i don t think z=7 0.9% vs 0.1%; not true z=7 0.6% vs 0.0%; no way z=7 0.9% vs 0.1%

2–4-word phrases, disgust: the worst z=18 6.2% vs 0.2%; is the worst z=10 1.6% vs 0.0%; is disgusting z=9 1.9% vs 0.0%; is awful z=8 1.2% vs 0.0%; worse than z=8 1.5% vs 0.1%; i've ever z=8 1.5% vs 0.1%; an awful z=7 1.0% vs 0.0%; even worse z=7 1.0% vs 0.0%; a weird z=7 1.2% vs 0.1%; i've ever seen z=7 1.0% vs 0.0%; weird to z=7 0.9% vs 0.0%; this is the worst z=7 0.7% vs 0.0%; it's weird z=6 0.7% vs 0.0%; ever seen z=6 1.2% vs 0.1%; i hate z=6 2.2% vs 0.4%; disgusting and z=6 0.7% vs 0.0%; that's awful z=6 0.6% vs 0.0%; be worse z=6 0.6% vs 0.0%; absolute worst z=6 0.6% vs 0.0%; is weird z=6 0.6% vs 0.0%

2–4-word phrases, embarrassment: a shame z=8 2.7% vs 0.1%; be ashamed z=7 1.5% vs 0.0%; my bad z=6 2.3% vs 0.1%; s a shame z=6 1.2% vs 0.0%; i wear z=6 1.2% vs 0.0%; ashamed of z=6 1.2% vs 0.0%; ashamed to z=6 1.2% vs 0.0%; of shame z=6 1.2% vs 0.0%; it s a shame z=6 1.2% vs 0.0%; an embarrassment z=5 1.5% vs 0.0%; my body z=5 1.2% vs 0.0%; sorry i z=5 2.3% vs 0.2%; second hand z=5 0.8% vs 0.0%; even know that z=5 0.8% vs 0.0%; it was weird z=5 0.8% vs 0.0%; body and z=5 0.8% vs 0.0%; was weird z=5 0.8% vs 0.0%; to much z=5 0.8% vs 0.0%; an easy z=5 0.8% vs 0.0%; be ashamed of z=5 0.8% vs 0.0%

2–4-word phrases, excitement: wait to z=14 3.2% vs 0.0%; can't wait z=13 3.0% vs 0.1%; t wait z=12 2.6% vs 0.0%; can t wait z=12 2.6% vs 0.0%; excited to z=12 2.7% vs 0.0%; wait for z=11 2.3% vs 0.1%; excited for z=11 3.3% vs 0.0%; happy new z=10 2.0% vs 0.1%; cake day z=10 2.0% vs 0.1%; so excited z=10 1.9% vs 0.0%; happy new year z=10 1.7% vs 0.1%; excited to see z=9 1.6% vs 0.0%; happy birthday z=9 1.4% vs 0.0%; happy cake day z=9 1.6% vs 0.1%; happy cake z=9 1.6% vs 0.1%; can t wait to z=9 1.4% vs 0.0%; t wait to z=9 1.4% vs 0.0%; wait to see z=9 1.3% vs 0.0%; i can't wait z=9 1.3% vs 0.0%; new year z=9 1.7% vs 0.1%

2–4-word phrases, fear: a terrible z=13 3.3% vs 0.1%; afraid of z=13 3.3% vs 0.0%; afraid to z=11 2.3% vs 0.0%; scared of z=10 2.5% vs 0.0%; i'm afraid z=10 2.6% vs 0.0%; a horrible z=9 1.9% vs 0.1%; scared to z=8 2.8% vs 0.0%; is terrible z=8 1.4% vs 0.0%; worried about z=8 1.6% vs 0.1%; fear of z=8 1.2% vs 0.0%; i'm scared z=7 1.8% vs 0.0%; terrible idea z=7 1.2% vs 0.0%; a terrible idea z=7 1.2% vs 0.0%; was afraid z=7 1.2% vs 0.0%; out of me z=7 0.9% vs 0.0%; be afraid z=6 0.9% vs 0.0%; afraid i z=6 1.1% vs 0.0%; is terrifying z=6 1.1% vs 0.0%; is scary z=6 0.9% vs 0.0%; more worried about z=6 0.7% vs 0.0%

2–4-word phrases, gratitude: thank you z=50 35.2% vs 0.3%; thanks for z=37 18.9% vs 0.2%; you for z=30 10.0% vs 0.2%; for the z=29 12.9% vs 1.4%; thank you for z=27 10.0% vs 0.1%; thanks for the z=24 8.9% vs 0.1%; you i z=18 3.6% vs 0.2%; for your z=17 3.7% vs 0.3%; thanks i z=17 3.1% vs 0.1%; you so z=16 2.9% vs 0.1%; thank you i z=16 3.5% vs 0.0%; for sharing z=15 2.7% vs 0.1%; you so much z=15 2.7% vs 0.0%; thank you so z=14 2.9% vs 0.0%; thank you so much z=14 2.7% vs 0.0%; so much z=13 3.7% vs 0.6%; thanks for sharing z=12 1.6% vs 0.0%; you for the z=12 2.2% vs 0.0%; you for your z=12 1.8% vs 0.0%; i will z=12 2.5% vs 0.3%

2–4-word phrases, grief: your loss z=8 7.5% vs 0.1%; sorry about your z=7 4.5% vs 0.0%; sorry for z=7 9.0% vs 0.2%; sorry about z=6 4.5% vs 0.0%; passed away z=6 3.0% vs 0.0%; my condolences z=6 3.0% vs 0.0%; your friend z=6 4.5% vs 0.1%; for your loss z=5 4.5% vs 0.1%; sorry for your loss z=5 4.5% vs 0.1%; he died z=5 3.0% vs 0.0%; very sorry z=5 3.0% vs 0.0%; my only z=5 3.0% vs 0.0%; sorry for your z=5 4.5% vs 0.1%; this may z=5 3.0% vs 0.0%; about your z=5 4.5% vs 0.1%; to death z=5 3.0% vs 0.0%; two years z=5 3.0% vs 0.0%; after he z=5 3.0% vs 0.0%; to die z=4 3.0% vs 0.1%; he could z=4 3.0% vs 0.1%

2–4-word phrases, joy: glad you z=17 4.0% vs 0.1%; so happy z=16 3.6% vs 0.1%; glad to z=15 3.1% vs 0.1%; happy to z=14 2.9% vs 0.1%; i'm glad z=14 3.1% vs 0.1%; happy for z=13 2.5% vs 0.1%; so glad z=13 2.3% vs 0.1%; glad i z=12 2.1% vs 0.1%; i enjoy z=11 1.9% vs 0.0%; cake day z=11 1.7% vs 0.1%; enjoy it z=11 1.6% vs 0.0%; happy cake day z=11 1.6% vs 0.0%; happy cake z=11 1.6% vs 0.0%; me happy z=10 1.5% vs 0.0%; glad to see z=10 1.5% vs 0.0%; i enjoyed z=10 1.7% vs 0.0%; be happy z=10 1.5% vs 0.1%; enjoy the z=10 1.3% vs 0.0%; happy for you z=10 1.3% vs 0.0%; m glad z=10 1.4% vs 0.1%

2–4-word phrases, love: i love z=54 36.9% vs 0.5%; love it z=26 8.0% vs 0.2%; i like z=25 8.2% vs 0.5%; love the z=23 6.2% vs 0.2%; love this z=22 6.1% vs 0.1%; love to z=21 5.1% vs 0.1%; i love it z=19 4.6% vs 0.1%; would love z=17 3.4% vs 0.1%; i love this z=17 4.1% vs 0.0%; i love the z=17 3.4% vs 0.1%; love that z=16 3.4% vs 0.0%; love you z=16 3.3% vs 0.0%; love how z=16 3.5% vs 0.0%; love name z=15 2.8% vs 0.1%; my favorite z=15 3.3% vs 0.2%; i love how z=14 2.9% vs 0.0%; would love to z=14 2.2% vs 0.1%; i'd love z=14 2.2% vs 0.0%; i loved z=14 2.2% vs 0.1%; i would love z=13 2.0% vs 0.0%

2–4-word phrases, nervousness: i'm worried z=7 3.1% vs 0.0%; worried that z=7 3.1% vs 0.0%; worried about z=7 3.8% vs 0.1%; worried about the z=6 2.3% vs 0.0%; i'm worried that z=6 2.3% vs 0.0%; worried for z=6 3.1% vs 0.0%; to pick z=6 2.3% vs 0.0%; very worried z=5 1.5% vs 0.0%; get older z=5 1.5% vs 0.0%; so don't z=5 1.5% vs 0.0%; be worried z=5 1.5% vs 0.0%; so worried z=5 1.5% vs 0.0%; i can feel z=5 1.5% vs 0.0%; more worried about z=5 1.5% vs 0.0%; i have this z=5 1.5% vs 0.0%; more worried z=5 1.5% vs 0.0%; me off z=5 1.5% vs 0.0%; worrying about z=5 1.5% vs 0.0%; can feel z=4 1.5% vs 0.0%; makes me z=4 4.6% vs 0.5%

2–4-word phrases, neutral: name and z=8 1.4% vs 0.8%; in the z=7 3.7% vs 3.0%; and name z=7 1.2% vs 0.7%; name is z=6 1.8% vs 1.3%; name name z=6 0.6% vs 0.2%; they are z=6 1.0% vs 0.6%; name and name z=6 0.8% vs 0.5%; if they z=6 0.6% vs 0.3%; with name z=6 0.6% vs 0.3%; because they z=6 0.4% vs 0.2%; they don't z=6 0.3% vs 0.1%; on the z=5 1.9% vs 1.5%; name was z=5 0.7% vs 0.4%; name got z=5 0.2% vs 0.0%; he was z=5 0.9% vs 0.6%; looks like z=5 0.8% vs 0.5%; they can z=5 0.3% vs 0.1%; because she z=5 0.2% vs 0.0%; does not z=5 0.2% vs 0.1%; on name z=5 0.2% vs 0.1%

2–4-word phrases, optimism: i hope z=36 18.7% vs 0.4%; hope you z=22 7.1% vs 0.2%; good luck z=18 5.5% vs 0.3%; i hope you z=18 4.7% vs 0.1%; really hope z=12 2.2% vs 0.0%; hope it z=12 2.1% vs 0.0%; hope that z=12 2.1% vs 0.0%; i really hope z=12 2.1% vs 0.0%; hope he z=12 2.0% vs 0.0%; hope this z=11 1.9% vs 0.0%; hope the z=11 1.8% vs 0.0%; hope they z=11 1.7% vs 0.0%; hope she z=9 1.3% vs 0.0%; hope we z=9 1.2% vs 0.0%; hope you're z=9 1.2% vs 0.0%; was hoping z=9 1.2% vs 0.0%; hope for z=9 1.2% vs 0.0%; hope your z=9 1.2% vs 0.0%; i hope it z=9 1.2% vs 0.0%; just hope z=9 1.2% vs 0.0%

2–4-word phrases, pride: proud of z=16 21.0% vs 0.1%; so proud z=13 13.6% vs 0.0%; so proud of z=10 8.6% vs 0.0%; proud of you z=10 8.6% vs 0.1%; be proud z=8 4.9% vs 0.0%; of you z=7 8.6% vs 0.2%; i'm so proud z=7 4.9% vs 0.0%; so proud of you z=7 3.7% vs 0.0%; should be proud z=6 3.7% vs 0.0%; i'm so proud of z=6 3.7% vs 0.0%; of you and z=6 2.5% vs 0.0%; m proud z=6 2.5% vs 0.0%; i m proud z=6 2.5% vs 0.0%; proud to z=6 2.5% vs 0.0%; i'm proud z=5 2.5% vs 0.0%; be proud of z=5 2.5% vs 0.0%; this community z=5 2.5% vs 0.0%; i am just z=5 2.5% vs 0.0%; been so z=5 2.5% vs 0.0%; am just z=5 2.5% vs 0.0%

2–4-word phrases, realization: i realized z=10 2.2% vs 0.0%; realize that z=9 1.6% vs 0.1%; i forgot z=9 1.6% vs 0.1%; i thought z=8 3.7% vs 0.6%; i didn't z=8 2.9% vs 0.4%; didn't realize z=8 1.2% vs 0.0%; reminds me z=8 1.6% vs 0.1%; you realize z=8 1.0% vs 0.0%; i didn't realize z=8 1.0% vs 0.0%; just realized z=8 1.0% vs 0.0%; reminds me of z=8 1.5% vs 0.1%; me of z=8 1.7% vs 0.1%; i was z=7 6.4% vs 1.8%; i thought it z=7 1.5% vs 0.1%; it took z=7 0.9% vs 0.0%; realized that z=6 0.8% vs 0.0%; realize it z=6 0.7% vs 0.0%; thought it z=6 1.6% vs 0.2%; to realize z=6 0.7% vs 0.0%; but then i z=6 0.7% vs 0.0%

2–4-word phrases, relief: thank god z=10 5.6% vs 0.0%; m glad z=9 5.6% vs 0.1%; i m glad z=9 5.6% vs 0.1%; glad i z=8 5.6% vs 0.1%; at least z=7 10.3% vs 0.7%; m glad i z=7 3.2% vs 0.0%; i m glad i z=7 3.2% vs 0.0%; have to worry z=6 2.4% vs 0.0%; have to worry about z=6 2.4% vs 0.0%; glad i'm not z=6 2.4% vs 0.0%; makes me feel z=6 3.2% vs 0.1%; glad i'm z=6 2.4% vs 0.0%; someone i z=6 2.4% vs 0.0%; this makes me feel z=6 2.4% vs 0.0%; thank name z=6 3.2% vs 0.1%; so glad z=6 4.0% vs 0.1%; better about z=6 2.4% vs 0.0%; the only one z=6 4.0% vs 0.1%; don't have to z=6 2.4% vs 0.0%; to worry about z=6 2.4% vs 0.0%

2–4-word phrases, remorse: sorry i z=20 10.1% vs 0.1%; so sorry z=20 10.1% vs 0.1%; i'm sorry z=19 9.3% vs 0.1%; sorry for z=19 8.6% vs 0.1%; i m sorry z=16 6.1% vs 0.1%; m sorry z=16 6.1% vs 0.1%; sorry but z=15 5.5% vs 0.1%; sorry you z=14 5.1% vs 0.1%; i m so sorry z=13 4.0% vs 0.0%; m so sorry z=13 4.0% vs 0.0%; sorry to z=12 3.8% vs 0.1%; m so z=12 4.2% vs 0.1%; i m so z=12 4.2% vs 0.1%; sorry that z=11 3.0% vs 0.1%; i'm so sorry z=10 2.7% vs 0.0%; sorry for the z=10 2.5% vs 0.0%; sorry for your z=10 2.5% vs 0.1%; i m z=10 11.4% vs 2.1%; i regret z=9 2.3% vs 0.0%; i meant z=9 2.5% vs 0.1%

2–4-word phrases, sadness: so sad z=12 2.1% vs 0.0%; i feel z=12 4.3% vs 0.6%; bad for z=12 2.0% vs 0.1%; i miss z=11 2.0% vs 0.1%; sorry for z=11 2.2% vs 0.2%; feel bad z=11 1.7% vs 0.1%; i'm sorry z=10 2.2% vs 0.2%; feel bad for z=10 1.5% vs 0.0%; sad that z=10 1.7% vs 0.0%; sorry to z=9 1.4% vs 0.1%; so sorry z=9 1.9% vs 0.2%; me sad z=9 1.3% vs 0.0%; i m sorry z=9 1.6% vs 0.1%; m sorry z=9 1.6% vs 0.1%; sorry for your z=9 1.2% vs 0.1%; sorry that z=9 1.2% vs 0.1%; i feel bad z=8 1.0% vs 0.0%; for your loss z=8 1.1% vs 0.0%; sorry for your loss z=8 1.1% vs 0.0%; your loss z=8 1.1% vs 0.1%

2–4-word phrases, surprise: i wonder z=18 4.7% vs 0.1%; wow i z=13 2.9% vs 0.1%; wonder if z=13 2.5% vs 0.0%; can't believe z=12 2.1% vs 0.1%; wonder how z=12 2.0% vs 0.0%; i wonder if z=11 1.8% vs 0.0%; oh my z=11 2.4% vs 0.2%; t believe z=11 1.7% vs 0.0%; was wondering z=10 1.7% vs 0.0%; be surprised z=10 1.8% vs 0.1%; can t believe z=9 1.4% vs 0.0%; wonder why z=9 1.3% vs 0.0%; wow you z=9 1.3% vs 0.0%; i wonder how z=9 1.2% vs 0.0%; surprised if z=9 1.2% vs 0.0%; a surprise z=9 1.2% vs 0.0%; i'm surprised z=8 2.5% vs 0.0%; i was wondering z=8 1.5% vs 0.0%; wonder what z=8 1.1% vs 0.0%; i can't believe z=8 1.1% vs 0.0%

Strong phrase flags (in ≥5% of one class's rows and at ≥4× the rate in the others):

remorse: sorry 75.9% vs 0.9%
love: love 69.0% vs 1.6%
gratitude: thanks 47.3% vs 0.5%
amusement: lol 42.6% vs 0.9%
desire: wish 39.7% vs 0.5%
pride: proud 38.3% vs 0.1%
gratitude: thank 38.2% vs 0.4%
love: i love 36.9% vs 0.5%
optimism: hope 36.0% vs 0.8%
gratitude: thank you 35.2% vs 0.3%
desire: i wish 29.9% vs 0.2%
grief: died 22.4% vs 0.1%
relief: glad 22.2% vs 1.0%
joy: happy 21.2% vs 0.7%
pride: proud of 21.0% vs 0.1%
gratitude: thanks for 18.9% vs 0.2%
optimism: i hope 18.7% vs 0.4%
sadness: sad 17.3% vs 0.2%
joy: glad 17.3% vs 0.6%
surprise: wow 17.0% vs 0.8%
anger: fuck 15.0% vs 0.5%
grief: sorry 14.9% vs 1.7%
desire: want 14.8% vs 1.7%
surprise: surprised 14.2% vs 0.1%
admiration: good 14.1% vs 2.9%
admiration: great 14.1% vs 0.8%
nervousness: worried 13.8% vs 0.1%
pride: so proud 13.6% vs 0.0%
disgust: disgusting 13.5% vs 0.0%
desire: wish i 13.1% vs 0.1%
gratitude: for the 12.9% vs 1.4%
sadness: sorry 12.8% vs 1.4%
confusion: or 12.5% vs 3.0%
desire: could 11.6% vs 1.5%
embarrassment: shame 11.5% vs 0.2%
embarrassment: embarrassing 11.5% vs 0.0%
nervousness: nervous 11.5% vs 0.0%
confusion: why 11.5% vs 2.0%
remorse: m 11.4% vs 2.1%
remorse: i m 11.4% vs 2.1%

Shortcut models

Predicting the label class on test (5,408 rows). Chance 3.6%, majority class ('neutral') 31.1%; balanced chance 3.6%.

model (logistic regression, trained on the train sample)	test accuracy	balanced accuracy (mean recall)
bag of words, whole state (words and word pairs)	53.2%	32.0%
bag of words, main text only (`comment`)	41.7%	20.5%
surface features only (no words: length, punctuation, case, markup, digits, script, state sizes, option count, instruction kind)	31.1%	5.0%
surface features of the main text only	31.0%	4.8%

Strongest single surface features (logistic regression on one feature, balanced accuracy on test):

feature	accuracy	balanced accuracy
ends_?	31.3%	5.2%
count_?	31.0%	3.8%
count_!	30.8%	3.6%
chars(log)	31.1%	3.6%
words(log)	31.1%	3.6%
upper_ratio	31.1%	3.6%
digit_ratio	31.1%	3.6%
nonlatin	31.1%	3.6%
emoji	31.1%	3.6%
html_tag	31.1%	3.6%

Other state fields alone (predicting the label class on test from one field, without the main text):

field	treated as	accuracy	balanced accuracy
source	categorical, 1 values	31.1%	3.6%

No-meaning option picker: a logistic ranker scores each option from its position, length, key type, fixed-option identity and shape (commas, brackets, capitals), never reading the state or the option's words, and picks the top option per row.

Test accuracy 31.1% against uniform chance 3.6% (this includes the fixed options, whose share is a class prior).

2. Duplicates and split separation

Families shared between splits

splits	shared families	examples
train ∩ dev	0
train ∩ calibration	0
train ∩ test	0
dev ∩ calibration	0
dev ∩ test	0
calibration ∩ test	0

Train rows whose main text repeats an earlier row's (normalised): 0 (0.0%); groups: 0; largest group 1.
Train rows identical in the whole prompt (state, options, instructions): 0.
Identical whole prompt, different answer: 0 groups (0 rows).
Identical main text, different label class: 0 groups (0 rows). (Can be legitimate when the rest of the state or the options differ.)

Main text of held-out rows found verbatim in train (normalised; the leak gate ignores short texts shared by many items):

split	rows	examples
dev	0 (0.0%)
calibration	0 (0.0%)
test	0 (0.0%)

Near duplicates (MinHash, word 3-gram Jaccard ≥ 0.8 on the main text)

Train: 39 near-duplicate pairs; 73 rows (0.2%) sit in 35 clusters; largest cluster 4; excess rows (cluster size − 1) 38 (0.1%).
Clusters with more than one label class: 15 (32 rows).
- ×4: "I'm sorry for your loss" → caring 1, grief 1, remorse 1, amusement 1
- ×2: "He's hot" → neutral 1, admiration 1
- ×2: "[NAME]. What a time to be alive." → neutral 1, approval 1
- ×2: "because it’s funny" → amusement 1, confusion 1
- ×2: "Oh that’s terrifying" → surprise 1, fear 1
Held-out rows with a near duplicate in train: dev 2 (0.2%), calibration 0 (0.0%), test 6 (0.1%)
- train "Weird flex but okay" ~ test "Weird flex but ok" (J=0.87)
- train "What? That doesn’t answer my question." ~ test "That doesn't answer my question." (J=0.80)
- train "Love the username" ~ test "Love the username <3" (J=0.87)
- train "Thanks a bunch!" ~ test "Thanks a bunch <3" (J=0.83)
- train "Yes YES #YES" ~ test "* yes * yes * yes * yes" (J=1.00)

Largest train clusters:

×4: "I'm sorry for your loss"
×3: "HAHAHAHA!!!!!"
×2: "He's hot"
×2: "RemindMe! 7 days"
×2: "RemindMe! 3 Days"

5. Junk

split	empty main text	main text under 10 characters
train	0	240
dev	0	5
calibration	0	6
test	0	40

Very short train examples: "He's hot" (neutral); "by [NAME]" (neutral); "Hey now!" (approval); "$390 CAD." (neutral); "[NAME] 3!" (neutral); "Go you!" (neutral); "Shame." (neutral); "Wow. Yes" (approval); "funny as." (amusement); "Or both!" (neutral); "Lol, wut" (amusement); "My bad." (neutral)

Pattern scan of train main texts (count, then the share of each class's rows):

pattern	rows	by class
placeholder [NAME]-style	6,405 (13.9%)	admiration 14.5%, amusement 13.0%, anger 13.9%, annoyance 12.3%, approval 11.4%, caring 9.2%, confusion 12.9%, curiosity 13.1%, desire 17.0%, disappointment 13.9%, disapproval 10.6%, disgust 11.4%, embarrassment 7.3%, excitement 13.0%, fear 13.5%, gratitude 7.9%, grief 14.9%, joy 9.3%, love 19.4%, nervousness 7.7%, neutral 17.0%, optimism 12.9%, pride 11.1%, realization 10.3%, relief 10.3%, remorse 9.7%, sadness 10.2%, surprise 15.9%
lorem ipsum	0 (0.0%)
TODO/TBD/FIXME	0 (0.0%)
'As an AI' / refusal	1 (0.0%)	annoyance 0.0%
chat preamble ('Here is/are...', 'Sure!')	39 (0.1%)	admiration 0.1%, approval 0.4%, curiosity 0.1%, disappointment 0.1%, disapproval 0.1%, excitement 0.3%, gratitude 0.0%, joy 0.1%, neutral 0.1%, optimism 0.1%, sadness 0.2%
meta words (example/variation/message:)	0 (0.0%)
model thinking tags	0 (0.0%)
JSON/code-fence leftovers	0 (0.0%)
encoding garbage (mojibake/replacement char)	0 (0.0%)
HTML tag	3 (0.0%)	approval 0.0%, neutral 0.0%
HTML entity	0 (0.0%)
base64-like run (40+ chars)	11 (0.0%)	admiration 0.0%, amusement 0.0%, anger 0.1%, confusion 0.1%, curiosity 0.1%, neutral 0.0%, surprise 0.1%
URL	0 (0.0%)

placeholder [NAME]-style: emotion:go_emotions:train:efeo8w6:emotion (admiration): Unreal. [NAME]- Expert level | emotion:go_emotions:train:eeezcc7:emotion (disapproval): [NAME] would've had to get lucky on that one. Can't give up that opportunity. | emotion:go_emotions:train:eeajm3e:emotion (approval): I bet he’s not going to visit their hall of fame in order to avoid meeting w/ [NAME]
'As an AI' / refusal: emotion:go_emotions:train:ef9ebro:emotion (annoyance): I cannot help but get a sense of schadenfreude when a seemingly shitty person ruins their own career right before my eyes. ☕️
chat preamble ('Here is/are...', 'Sure!'): emotion:go_emotions:train:eebqp3y:emotion (neutral): Sure, there are edge cases where this is less clear, but in the vast majority of conflict, this is true. | emotion:go_emotions:train:ee7fxah:emotion (neutral): Sure, its just that it was completely irrelevant to my reply | emotion:go_emotions:train:ed1m9qs:emotion (approval): Sure, they can help nurse [NAME] back to health.
HTML tag: emotion:go_emotions:train:ee5v5uy:emotion (neutral): <<creepy loui face.jpg>> | emotion:go_emotions:train:edfd4z0:emotion (approval): Yes, definitely classy. <rolls eyes> | emotion:go_emotions:train:ef8b1z0:emotion (neutral): If by "better", you mean "more fun", then I'd go with the second one. <smile>
base64-like run (40+ chars): emotion:go_emotions:train:ef9cbbq:emotion (surprise): Oh no, my secret identity as a neurosurgeon/entomologist/radiologist/anesthetist/dentist has been discovered! I knew I should have given le… | emotion:go_emotions:train:edrtk8u:emotion (amusement): I did anyway hahahahahahahahahahahahahahahahahahahahahahahahahahhahahahhahahaa fucking rekt XD | emotion:go_emotions:train:ef59tsi:emotion (neutral): Omg ur like soooooooooooooooooooooooooooooooooooooooooooooooooo quirky!
Possibly cut off: 1 of 4 train main texts over 300 characters end mid-sentence (letter, digit or comma). By class: admiration 100.0%, neutral 0.0%
- emotion:go_emotions:train:edy8hly:emotion: …765434567654323454323456543345678987654323456789876565656565656565656565656565454545654565454323456765432345678765456 IQ

Repeated sentences across rows (≥25 characters, in at least 0.2% of the sample):

none

6. Samples

20 random train rows per kind: emotion-samples.txt. Reading notes are in the findings above.