G-reen/readme_card_preview_1

FastDetector · Stats ReleaseG-reen/cc-2021-stat · Common Crawl CC-MAIN-2021-49

Can a detector tell
the rewrite from the original?

23,214 human-written web documents, each paired with an AI-generated counterpart produced by one of seven open-weight generator configurations under four prompt types, then measured with ten text-distance metrics and scored by thirteen AI-text detectors.

Human text — original   AI text — final_response

23,214

human / AI text pairs

7

generator configs

4

prompt types · 118 templates

13

detectors scored

0.8775

best AUROC

01Findings02Anatomy of a pair03Detector leaderboard04Where detectors fail05How far the text moved06Detector dossiers07Statistics08Using the data09Method & provenance

■  01 — Findings

What the numbers say

The headline numbers, read off the evaluation split.

Strongest detector

EditLens Roberta-Large Score

0.8775AUROC

At its validation-chosen threshold (0.5961, target FPR ≤ 0.5%) it flags 0.3588 of AI texts at an FPR of 0.0046.

Hardest prompt type

rewrite

0.0536avg TPR

Mean true-positive rate across all 13 detectors (mean AUROC 0.5314). Its instructions are constrained edits that keep the rest of the document unchanged or verbatim.

Hardest generator config

Gemma 4 E4B it · T=0.7

0.0693avg TPR

gemma-4-E4B-it (Temp: 0.7) — mean TPR across all 13 detectors (mean AUROC 0.5472).

Two tiers. The two EditLens RoBERTa-Large outputs are the only detectors above 0.8 AUROC overall (0.8775 and 0.8145). The other eleven (perplexity, entropy, top-p and top-k outliers and FastDetectGPT under two Llama-3.2-3B checkpoints, plus Binoculars) span 0.4854 to 0.6302 AUROC, and two of them fall below the 0.5 chance line.

■  02 — Anatomy of a pair

Anatomy of a pair

Every row is one human document and one AI text derived from it. This is how each pair was made.

STEP 1

Human document

English web text from CC-MAIN-2021-49 via G-reen/cc-2021-raw, sharded and filtered

STEP 2

Prompt

One of 118 templates in four prompt types wraps the document, in one or two turns

STEP 3

Generator

One of seven model / sampling configs answers each turn

STEP 4

AI text

The last turn's response becomes final_response

STEP 5

Scoring

10 distance metrics per pair, 13 detector scores per side

Four prompt types, 118 templates

direct_reference5,735 rows

3 templates · 3 one-turn (5,735 rows)

Write a new text that references the original: same theme, borrowed style, or reused phrasing.

Instructions used

• Write another piece of text that is of a similar theme or background as this text.
• Take inspiration from the style, language, and content of this text, and write a new piece of text on a topic of your choice. However, you should write the new text as if it was a rough draft being typed out on a phone.
• Write a new piece of text that is on a completely different topic, but reuses verbatim as many phrases and sentences from this text as possible.

Most common template, verbatim (2,019 rows)

TURN 1

<document>
{{DOC}}
</document>

Write another piece of text that is of a similar theme or background as this text.
Output the full new text with no extra statements or commentations.
Begin directly with the text itself. Do not add a title, a heading, or a label naming what you have written.

indirect_reference5,653 rows

5 templates · 5 two-turn (5,653 rows)

Turn 1 converts the original into an intermediate form; turn 2 produces the final text from that output.

Instructions used

• Translate the given text to Chinese. → then: Translate this text to English. Output only the English translation with no extra commentary.
• Provide only the first and last word of every sentence in this text, replacing the middle of each sentence with an ellipsis '...'. Maintain the original paragraph structure. → then: Above is a partial trace of a human written document, in some arbitrary format. Based on that descriptor, recreate the original human writte…
• Formulate a set of strict, absolute constraints regarding format, word count, vocabulary restrictions, and syntax rules that, if followed perfectly by an AI, would result in generating a text nearly identical to this one. Note that your response should be formatted as a prompt that would be sent to an AI requesting it to generate something, not just a list of rules. → then: Output only the generated text with no extra commentary.
• Write a detailed descriptor for the manner in which the author of this uses language, and things such as their word choice. → then: Above is an AI generated descriptor/trace of some human written document, in some arbitrary format. Based on that descriptor, recreate the o…
• Write a detailed descriptor for the style of writing of this text, including its qualitative and quantitative properties. → then: Above is an AI generated descriptor/trace of some human written document, in some arbitrary format. Based on that descriptor, recreate the o…

Most common template, verbatim (1,470 rows)

TURN 1

<document>
{{DOC}}
</document>

Translate the given text to Chinese.
Do not output anything besides what you were requested to write, and do not output any extra commentary.

TURN 2

{{RESP_0}}
Translate this text to English. Output only the English translation with no extra commentary.
Begin directly with the text itself. Do not add a title, a heading, or a label naming what you have written.

revise5,940 rows

68 templates · 8 one-turn (3,119 rows) · 60 two-turn (2,821 rows)

Holistic edits: tone, audience, clarity, concision, expansion or structure. Two-turn variants chain two edits.

Instructions used

• Expand this text by adding backstory and context to enrich understanding.
• Restructure this text to improve the transitions and connections between ideas.
• Clarify this text by paraphrasing it with different vocabulary and varied sentence structures.
• Edit this text to eliminate redundant and filler words.
• Change the tone of this text to be more empathetic and understanding.
• Adapt this text for a lay audience with no technical background.
• Change the tone of this text to be more suited for a business audience.
• Change the tone of this text to be more relaxed and friendly.

Most common template, verbatim (479 rows)

TURN 1

<document>
{{DOC}}
</document>

Expand this text by adding backstory and context to enrich understanding.
Output the full new text with no extra statements or commentations.
Begin directly with the text itself. Do not add a title, a heading, or a label naming what you have written.

rewrite5,886 rows

42 templates · 6 one-turn (3,069 rows) · 36 two-turn (2,817 rows)

Constrained edits that change a specific part and keep the rest unchanged or verbatim. Two-turn variants chain two edits.

Instructions used

• Edit this text by replacing, removing or inserting twenty to thirty words, without changing anything else.
• Edit this text by combining at least four pairs of short, adjacent sentences into complex sentences using appropriate conjunctions, leaving all other sentences untouched.
• Edit this text by completely rewriting the topic sentence of every paragraph, while preserving the supporting sentences in those paragraphs verbatim.
• Edit this text by removing all adverbs, making only minimal adjustments to the surrounding grammar to ensure the sentences remain structurally sound.
• Edit this text by removing one section from it (the least necessary), you may only make minimal modifications to adjacent sections for better flow.
• Modify this text to be as long as possible. At least half of the original text must be repeated verbatim, and at most three quarters of it; the remainder must be new prose.

Most common template, verbatim (557 rows)

TURN 1

<document>
{{DOC}}
</document>

Edit this text by replacing, removing or inserting twenty to thirty words, without changing anything else.
Output the full new text with no extra statements or commentations.
Begin directly with the text itself. Do not add a title, a heading, or a label naming what you have written.

Seven generator configurations

Each dataset config (shard) holds exactly one generator configuration. All use presence penalty 1.5. Row counts cover the full dataset.

Generator	Config	Sampling	Rows
Granite 4.1 8B cyankiwi/granite-4.1-8b-AWQ-INT4	shard_0	T 0.6 · top_p 0.9 · top_k 40	3,275
Llama 3.1 8B Instruct hugging-quants/Meta-Llama-3.1-8B-Instruct-AWQ-INT4	shard_1	T 0.6 · top_p 0.9 · top_k 40	3,327
Qwen3 8B Qwen/Qwen3-8B-AWQ	shard_2	T 0.7 · top_p 0.8 · top_k 20 · thinking off	3,391
Gemma 4 E4B it google/gemma-4-E4B-it	shard_3	T 0.7 · top_p 0.8 · top_k -1 · thinking off	3,323
Ministral 3 8B Instruct cyankiwi/Ministral-3-8B-Instruct-2512-AWQ-4bit	shard_4	T 0.6 · top_p 0.9 · top_k 40	3,258
Llama 3.1 8B Instruct hugging-quants/Meta-Llama-3.1-8B-Instruct-AWQ-INT4	shard_5	T 1.25 · top_p 1.0 · top_k -1 · top_a 0.1 · nsigma 1.5	3,325
Granite 4.1 8B cyankiwi/granite-4.1-8b-AWQ-INT4	shard_6	T 1.25 · top_p 1.0 · top_k -1 · top_a 0.1 · nsigma 1.5	3,315

■  03 — Detector leaderboard

Detector leaderboard

All thirteen detectors on the full evaluation split, ranked by AUROC.

Detector · rule	AUROC	Threshold	TPR	FPR	Acc	F1
01EditLens Roberta-Large Score higher → AI · FPR ≤ 0.5%	0.8775	0.5961	0.3588	0.0046	0.6771	0.5263
02EditLens Roberta-Large Bucket higher → AI · best F1	0.8145	0.0000	0.6656	0.0513	0.8072	0.7754
03Binoculars higher → AI · FPR ≤ 0.5%	0.6302	1.0472	0.0121	0.0031	0.5045	0.0239
04FastDetectGPT (Llama-3.2-3B-Instruct) higher → AI · FPR ≤ 0.5%	0.5725	6.7770	0.0027	0.0041	0.4993	0.0054
05Top-k Outliers (Llama-3.2-3B-Instruct) lower → AI · FPR ≤ 0.5%	0.5637	0.0109	0.0191	0.0033	0.5079	0.0373
06Top-p Outliers (Llama-3.2-3B-Instruct) lower → AI · FPR ≤ 0.5%	0.5598	0.0158	0.0108	0.0030	0.5039	0.0213
07Perplexity (Llama-3.2-3B-Instruct) lower → AI · FPR ≤ 0.5%	0.5478	2.2698	0.0055	0.0045	0.5005	0.0108
08Entropy (Llama-3.2-3B-Instruct) lower → AI · FPR ≤ 0.5%	0.5391	0.7575	0.0032	0.0044	0.4994	0.0063
09Top-p Outliers (Llama-3.2-3B) lower → AI · FPR ≤ 0.5%	0.5168	0.0054	0.0062	0.0071	0.4995	0.0122
10Top-k Outliers (Llama-3.2-3B) lower → AI · FPR ≤ 0.5%	0.5168	0.0056	0.0053	0.0048	0.5002	0.0104
11FastDetectGPT (Llama-3.2-3B) lower → AI · FPR ≤ 0.5%	0.5159	-3.2326	0.0986	0.0066	0.5460	0.1784
12Entropy (Llama-3.2-3B) lower → AI · FPR ≤ 0.5%	0.4869	0.2888	0.0005	0.0049	0.4978	0.0010
13Perplexity (Llama-3.2-3B) lower → AI · FPR ≤ 0.5%	0.4854	1.2916	0.0005	0.0044	0.4980	0.0010

Evaluation split, n = 41,784 scored texts (20,892 human + 20,892 AI). Bars run from 0 to 1 and the tick marks 0.5 (chance). Each threshold was chosen on the 2,322-row validation split for the target shown under its name, then applied unchanged to the evaluation split. Select a detector name to open its dossier.

■  04 — Where detectors fail

Where detectors fail

The same detectors, broken down by prompt type and by generator config.

		PROMPT TYPE				GENERATOR CONFIG
Detector	All	direct	indirect	revise	rewrite	Llama 0.6	Llama 1.25	Ministral 0.6	Qwen3 0.7	Gemma 0.7	Granite 0.6	Granite 1.25
EditLens Score	0.877	0.841	0.933	0.966	0.771	0.818	0.854	0.991	0.808	0.852	0.903	0.925
EditLens Bucket	0.815	0.790	0.852	0.924	0.693	0.709	0.771	0.974	0.734	0.785	0.853	0.884
Binoculars	0.630	0.601	0.650	0.665	0.605	0.531	0.882	0.543	0.484	0.647	0.607	0.714
FastDetectGPT · Inst	0.572	0.590	0.630	0.540	0.533	0.694	0.542	0.592	0.589	0.569	0.544	0.471
Top-k Outliers · Inst	0.564	0.616	0.566	0.600	0.472	0.720	0.259	0.642	0.691	0.505	0.605	0.523
Top-p Outliers · Inst	0.560	0.584	0.598	0.532	0.528	0.696	0.570	0.556	0.615	0.512	0.530	0.433
Perplexity · Inst	0.548	0.578	0.559	0.587	0.467	0.739	0.211	0.608	0.687	0.475	0.614	0.499
Entropy · Inst	0.539	0.567	0.541	0.580	0.468	0.703	0.247	0.574	0.665	0.461	0.607	0.517
Top-p Outliers · Base	0.517	0.611	0.525	0.495	0.440	0.675	0.220	0.601	0.666	0.485	0.559	0.410
Top-k Outliers · Base	0.517	0.562	0.511	0.546	0.447	0.678	0.210	0.586	0.666	0.451	0.557	0.468
FastDetectGPT · Base	0.516	0.429	0.483	0.554	0.594	0.294	0.899	0.398	0.308	0.552	0.468	0.692
Entropy · Base	0.487	0.476	0.478	0.534	0.457	0.638	0.215	0.497	0.611	0.414	0.552	0.481
Perplexity · Base	0.485	0.507	0.485	0.515	0.434	0.677	0.156	0.533	0.648	0.406	0.551	0.427

0.10.9AUROC on the evaluation split; hover a cell for its TPR and n. Column heads give generator family and temperature. Grey is chance (0.5); teal is better than chance; red is worse than chance, meaning the detector ranks human text as more AI-like. Inst = Llama-3.2-3B-Instruct, Base = Llama-3.2-3B.

Averaged over all 13 detectors

Each metric is the unweighted mean across the 13 detectors within one subset of the evaluation split, sorted by mean AUROC.

Prompt type	Avg AUROC	Avg TPR	Avg FPR	Avg Acc	Avg F1
revise	0.6184	0.1120	0.0082	0.5519	0.1353
indirect_reference	0.6009	0.1086	0.0076	0.5505	0.1445
direct_reference	0.5961	0.0921	0.0086	0.5417	0.1263
rewrite	0.5314	0.0536	0.0081	0.5227	0.0812

Generator config	Avg AUROC	Avg TPR	Avg FPR	Avg Acc	Avg F1
Llama 3.1 8B Instruct T=0.6	0.6594	0.0700	0.0074	0.5313	0.1065
Qwen3 8B T=0.7	0.6285	0.0704	0.0077	0.5313	0.1015
Ministral 3 8B Instruct T=0.6	0.6228	0.1264	0.0086	0.5589	0.1383
Granite 4.1 8B T=0.6	0.6115	0.0940	0.0086	0.5427	0.1174
Granite 4.1 8B T=1.25	0.5726	0.1051	0.0083	0.5484	0.1302
Gemma 4 E4B it T=0.7	0.5472	0.0693	0.0069	0.5312	0.0935
Llama 3.1 8B Instruct T=1.25	0.4643	0.1065	0.0097	0.5484	0.1412

■  05 — How far the text moved

How far the text moved

How different is each AI text from the document it came from? Ten metrics, from token overlap to embeddings and a reranker.

Headline views: jaccard_1 and cosdist between each human document and its AI counterpart, split by prompt type and by generator config. Each row is a histogram of one subset on a shared count scale; orange ticks mark the subset median.

All ten distance metrics · overall histograms

Evaluation split. Where a panel says so, values beyond the 0.5th–99.5th percentile are pinned into the edge bins so the bulk of the distribution stays readable.

All ten distance metrics · by prompt type

All ten distance metrics · by generator config

■  06 — Detector dossiers

Detector dossiers

One panel per detector: thresholding, score distributions, the validation sweep and per-subset performance.

01EditLens Roberta-Large ScoreAUROC 0.8775

DIRECTION

higher_is_ai

SWEPT FOR

fpr_0_5pct

THRESHOLD

0.5961

OVERALL AUROC

0.8775

Candidate operating points found on the validation split (the one used is highlighted):

best accuracy = 0.1533 best F1 = 0.1134 FPR ≤ 1% = 0.4161 FPR ≤ 0.5% = 0.5961 FPR ≤ 0.1% = 0.8442 FPR ≤ 0.01% = 0.9985

Subset	N	AUROC	TPR	FPR	Accuracy	F1
Overall	41,784	0.8775	0.3588	0.0046	0.6771	0.5263
Prompt  direct_reference	10,356	0.8405	0.4181	0.0041	0.7070	0.5880
Prompt  indirect_reference	10,182	0.9333	0.4331	0.0055	0.7138	0.6021
Prompt  revise	10,666	0.9662	0.4568	0.0039	0.7264	0.6254
Prompt  rewrite worst	10,580	0.7711	0.1304	0.0049	0.5628	0.2298
Generator  Llama 3.1 8B Instruct T=0.6	6,064	0.8177	0.2642	0.0040	0.6301	0.4166
Generator  Llama 3.1 8B Instruct T=1.25	6,004	0.8537	0.1666	0.0070	0.5798	0.2838
Generator  Ministral 3 8B Instruct T=0.6 best	5,860	0.9907	0.6451	0.0034	0.8208	0.7826
Generator  Qwen3 8B T=0.7	6,118	0.8077	0.2958	0.0072	0.6443	0.4541
Generator  Gemma 4 E4B it T=0.7	5,972	0.8523	0.2515	0.0027	0.6244	0.4011
Generator  Granite 4.1 8B T=0.6	5,876	0.9029	0.4364	0.0024	0.7170	0.6066
Generator  Granite 4.1 8B T=1.25	5,890	0.9249	0.4642	0.0054	0.7294	0.6317

Score histograms per prompt type (panels 1–4) and generator config (panels 5–11), evaluation split. The dashed line is the threshold. Select any chart to open it at full size.

02EditLens Roberta-Large BucketAUROC 0.8145

DIRECTION

higher_is_ai

SWEPT FOR

f1

THRESHOLD

0.0000

OVERALL AUROC

0.8145

Candidate operating points found on the validation split (the one used is highlighted):

best accuracy = 0 best F1 = 0 FPR ≤ 1% = 1 FPR ≤ 0.5% = 1 FPR ≤ 0.1% = 3 FPR ≤ 0.01% = 3

Subset	N	AUROC	TPR	FPR	Accuracy	F1
Overall	41,784	0.8145	0.6656	0.0513	0.8072	0.7754
Prompt  direct_reference	10,356	0.7897	0.6124	0.0508	0.7808	0.7364
Prompt  indirect_reference	10,182	0.8516	0.7362	0.0517	0.8423	0.8236
Prompt  revise	10,666	0.9239	0.8785	0.0491	0.9147	0.9115
Prompt  rewrite worst	10,580	0.6928	0.4352	0.0537	0.6907	0.5846
Generator  Llama 3.1 8B Instruct T=0.6	6,064	0.7088	0.4571	0.0508	0.7032	0.6063
Generator  Llama 3.1 8B Instruct T=1.25	6,004	0.7709	0.5936	0.0566	0.7685	0.7194
Generator  Ministral 3 8B Instruct T=0.6 best	5,860	0.9743	0.9703	0.0495	0.9604	0.9608
Generator  Qwen3 8B T=0.7	6,118	0.7343	0.5093	0.0517	0.7288	0.6526
Generator  Gemma 4 E4B it T=0.7	5,972	0.7853	0.6102	0.0492	0.7805	0.7354
Generator  Granite 4.1 8B T=0.6	5,876	0.8527	0.7372	0.0524	0.8424	0.8239
Generator  Granite 4.1 8B T=1.25	5,890	0.8838	0.7976	0.0489	0.8744	0.8639

Score histograms per prompt type (panels 1–4) and generator config (panels 5–11), evaluation split. The dashed line is the threshold. Select any chart to open it at full size.

03BinocularsAUROC 0.6302

DIRECTION

higher_is_ai

SWEPT FOR

fpr_0_5pct

THRESHOLD

1.0472

OVERALL AUROC

0.6302

Candidate operating points found on the validation split (the one used is highlighted):

best accuracy = 0.8856 best F1 = 0.7261 FPR ≤ 1% = 0.9974 FPR ≤ 0.5% = 1.047 FPR ≤ 0.1% = 1.09 FPR ≤ 0.01% = 1.161

Subset	N	AUROC	TPR	FPR	Accuracy	F1
Overall	41,784	0.6302	0.0121	0.0031	0.5045	0.0239
Prompt  direct_reference	10,356	0.6006	0.0114	0.0027	0.5043	0.0225
Prompt  indirect_reference	10,182	0.6500	0.0192	0.0037	0.5078	0.0376
Prompt  revise	10,666	0.6654	0.0075	0.0034	0.5021	0.0148
Prompt  rewrite	10,580	0.6046	0.0106	0.0026	0.5040	0.0209
Generator  Llama 3.1 8B Instruct T=0.6	6,064	0.5311	0.0013	0.0036	0.4988	0.0026
Generator  Llama 3.1 8B Instruct T=1.25 best	6,004	0.8816	0.0403	0.0033	0.5185	0.0772
Generator  Ministral 3 8B Instruct T=0.6	5,860	0.5429	0.0092	0.0038	0.5027	0.0182
Generator  Qwen3 8B T=0.7 worst	6,118	0.4838	0.0013	0.0036	0.4989	0.0026
Generator  Gemma 4 E4B it T=0.7	5,972	0.6467	0.0074	0.0033	0.5020	0.0146
Generator  Granite 4.1 8B T=0.6	5,876	0.6074	0.0065	0.0017	0.5024	0.0128
Generator  Granite 4.1 8B T=1.25	5,890	0.7140	0.0190	0.0024	0.5083	0.0372

Score histograms per prompt type (panels 1–4) and generator config (panels 5–11), evaluation split. The dashed line is the threshold. Select any chart to open it at full size.

04FastDetectGPT (Llama-3.2-3B-Instruct)AUROC 0.5725

DIRECTION

higher_is_ai

SWEPT FOR

fpr_0_5pct

THRESHOLD

6.7770

OVERALL AUROC

0.5725

Candidate operating points found on the validation split (the one used is highlighted):

best accuracy = -1.399 best F1 = -5.725 FPR ≤ 1% = 4.431 FPR ≤ 0.5% = 6.777 FPR ≤ 0.1% = 12.98 FPR ≤ 0.01% = 14.32

Subset	N	AUROC	TPR	FPR	Accuracy	F1
Overall	41,784	0.5725	0.0027	0.0041	0.4993	0.0054
Prompt  direct_reference	10,356	0.5900	0.0023	0.0048	0.4987	0.0046
Prompt  indirect_reference	10,182	0.6299	0.0020	0.0045	0.4987	0.0039
Prompt  revise	10,666	0.5402	0.0034	0.0047	0.4993	0.0067
Prompt  rewrite	10,580	0.5334	0.0032	0.0025	0.5004	0.0064
Generator  Llama 3.1 8B Instruct T=0.6 best	6,064	0.6945	0.0013	0.0053	0.4980	0.0026
Generator  Llama 3.1 8B Instruct T=1.25	6,004	0.5419	0.0080	0.0067	0.5007	0.0158
Generator  Ministral 3 8B Instruct T=0.6	5,860	0.5923	0.0007	0.0027	0.4990	0.0014
Generator  Qwen3 8B T=0.7	6,118	0.5891	0.0010	0.0016	0.4997	0.0020
Generator  Gemma 4 E4B it T=0.7	5,972	0.5690	0.0044	0.0057	0.4993	0.0086
Generator  Granite 4.1 8B T=0.6	5,876	0.5443	0.0017	0.0044	0.4986	0.0034
Generator  Granite 4.1 8B T=1.25 worst	5,890	0.4715	0.0020	0.0024	0.4998	0.0041

Score histograms per prompt type (panels 1–4) and generator config (panels 5–11), evaluation split. The dashed line is the threshold. Select any chart to open it at full size.

05Top-k Outliers (Llama-3.2-3B-Instruct)AUROC 0.5637

DIRECTION

lower_is_ai

SWEPT FOR

fpr_0_5pct

THRESHOLD

0.0109

OVERALL AUROC

0.5637

Candidate operating points found on the validation split (the one used is highlighted):

best accuracy = 0.06237 best F1 = 0.8383 FPR ≤ 1% = 0.02091 FPR ≤ 0.5% = 0.01089 FPR ≤ 0.1% = -4.941e-324 FPR ≤ 0.01% = -4.941e-324

Subset	N	AUROC	TPR	FPR	Accuracy	F1
Overall	41,784	0.5637	0.0191	0.0033	0.5079	0.0373
Prompt  direct_reference	10,356	0.6161	0.0373	0.0041	0.5166	0.0716
Prompt  indirect_reference	10,182	0.5662	0.0346	0.0018	0.5164	0.0667
Prompt  revise	10,666	0.5998	0.0038	0.0034	0.5002	0.0074
Prompt  rewrite	10,580	0.4725	0.0017	0.0038	0.4990	0.0034
Generator  Llama 3.1 8B Instruct T=0.6 best	6,064	0.7196	0.0627	0.0020	0.5303	0.1177
Generator  Llama 3.1 8B Instruct T=1.25 worst	6,004	0.2595	0.0013	0.0053	0.4980	0.0026
Generator  Ministral 3 8B Instruct T=0.6	5,860	0.6421	0.0010	0.0048	0.4981	0.0020
Generator  Qwen3 8B T=0.7	6,118	0.6911	0.0487	0.0023	0.5232	0.0927
Generator  Gemma 4 E4B it T=0.7	5,972	0.5050	0.0013	0.0013	0.5000	0.0027
Generator  Granite 4.1 8B T=0.6	5,876	0.6051	0.0123	0.0044	0.5039	0.0241
Generator  Granite 4.1 8B T=1.25	5,890	0.5226	0.0041	0.0027	0.5007	0.0081

Score histograms per prompt type (panels 1–4) and generator config (panels 5–11), evaluation split. The dashed line is the threshold. Select any chart to open it at full size.

06Top-p Outliers (Llama-3.2-3B-Instruct)AUROC 0.5598

DIRECTION

lower_is_ai

SWEPT FOR

fpr_0_5pct

THRESHOLD

0.0158

OVERALL AUROC

0.5598

Candidate operating points found on the validation split (the one used is highlighted):

best accuracy = 0.04982 best F1 = 0.1258 FPR ≤ 1% = 0.02116 FPR ≤ 0.5% = 0.01579 FPR ≤ 0.1% = 0.003676 FPR ≤ 0.01% = -4.941e-324

Subset	N	AUROC	TPR	FPR	Accuracy	F1
Overall	41,784	0.5598	0.0108	0.0030	0.5039	0.0213
Prompt  direct_reference	10,356	0.5837	0.0160	0.0029	0.5066	0.0315
Prompt  indirect_reference	10,182	0.5977	0.0246	0.0029	0.5108	0.0478
Prompt  revise	10,666	0.5320	0.0017	0.0036	0.4991	0.0034
Prompt  rewrite	10,580	0.5280	0.0017	0.0026	0.4995	0.0034
Generator  Llama 3.1 8B Instruct T=0.6 best	6,064	0.6963	0.0501	0.0030	0.5236	0.0952
Generator  Llama 3.1 8B Instruct T=1.25	6,004	0.5704	0.0043	0.0040	0.5002	0.0086
Generator  Ministral 3 8B Instruct T=0.6	5,860	0.5563	0.0010	0.0027	0.4991	0.0020
Generator  Qwen3 8B T=0.7	6,118	0.6147	0.0111	0.0026	0.5042	0.0219
Generator  Gemma 4 E4B it T=0.7	5,972	0.5118	0.0044	0.0020	0.5012	0.0087
Generator  Granite 4.1 8B T=0.6	5,876	0.5298	0.0027	0.0024	0.5002	0.0054
Generator  Granite 4.1 8B T=1.25 worst	5,890	0.4325	0.0010	0.0044	0.4983	0.0020

Score histograms per prompt type (panels 1–4) and generator config (panels 5–11), evaluation split. The dashed line is the threshold. Select any chart to open it at full size.

07Perplexity (Llama-3.2-3B-Instruct)AUROC 0.5478

DIRECTION

lower_is_ai

SWEPT FOR

fpr_0_5pct

THRESHOLD

2.2698

OVERALL AUROC

0.5478

Candidate operating points found on the validation split (the one used is highlighted):

best accuracy = 9.58 best F1 = 6608 FPR ≤ 1% = 2.862 FPR ≤ 0.5% = 2.27 FPR ≤ 0.1% = 1.429 FPR ≤ 0.01% = 1.201

Subset	N	AUROC	TPR	FPR	Accuracy	F1
Overall	41,784	0.5478	0.0055	0.0045	0.5005	0.0108
Prompt  direct_reference	10,356	0.5778	0.0062	0.0052	0.5005	0.0122
Prompt  indirect_reference	10,182	0.5590	0.0132	0.0027	0.5052	0.0259
Prompt  revise	10,666	0.5867	0.0013	0.0051	0.4981	0.0026
Prompt  rewrite	10,580	0.4672	0.0015	0.0047	0.4984	0.0030
Generator  Llama 3.1 8B Instruct T=0.6 best	6,064	0.7393	0.0254	0.0030	0.5112	0.0494
Generator  Llama 3.1 8B Instruct T=1.25 worst	6,004	0.2114	0.0007	0.0057	0.4975	0.0013
Generator  Ministral 3 8B Instruct T=0.6	5,860	0.6083	0.0007	0.0065	0.4971	0.0014
Generator  Qwen3 8B T=0.7	6,118	0.6866	0.0046	0.0033	0.5007	0.0091
Generator  Gemma 4 E4B it T=0.7	5,972	0.4750	0.0020	0.0033	0.4993	0.0040
Generator  Granite 4.1 8B T=0.6	5,876	0.6144	0.0037	0.0051	0.4993	0.0074
Generator  Granite 4.1 8B T=1.25	5,890	0.4990	0.0007	0.0044	0.4981	0.0014

Score histograms per prompt type (panels 1–4) and generator config (panels 5–11), evaluation split. The dashed line is the threshold. Select any chart to open it at full size.

08Entropy (Llama-3.2-3B-Instruct)AUROC 0.5391

DIRECTION

lower_is_ai

SWEPT FOR

fpr_0_5pct

THRESHOLD

0.7575

OVERALL AUROC

0.5391

Candidate operating points found on the validation split (the one used is highlighted):

best accuracy = 2.139 best F1 = 8.913 FPR ≤ 1% = 0.9471 FPR ≤ 0.5% = 0.7575 FPR ≤ 0.1% = 0.3863 FPR ≤ 0.01% = 0.2414

Subset	N	AUROC	TPR	FPR	Accuracy	F1
Overall	41,784	0.5391	0.0032	0.0044	0.4994	0.0063
Prompt  direct_reference	10,356	0.5668	0.0029	0.0050	0.4989	0.0057
Prompt  indirect_reference	10,182	0.5409	0.0077	0.0026	0.5026	0.0152
Prompt  revise	10,666	0.5799	0.0009	0.0047	0.4981	0.0019
Prompt  rewrite	10,580	0.4678	0.0013	0.0051	0.4981	0.0026
Generator  Llama 3.1 8B Instruct T=0.6 best	6,064	0.7033	0.0119	0.0026	0.5046	0.0234
Generator  Llama 3.1 8B Instruct T=1.25 worst	6,004	0.2467	0.0003	0.0053	0.4975	0.0007
Generator  Ministral 3 8B Instruct T=0.6	5,860	0.5743	0.0007	0.0058	0.4974	0.0014
Generator  Qwen3 8B T=0.7	6,118	0.6652	0.0039	0.0033	0.5003	0.0078
Generator  Gemma 4 E4B it T=0.7	5,972	0.4608	0.0017	0.0033	0.4992	0.0033
Generator  Granite 4.1 8B T=0.6	5,876	0.6065	0.0024	0.0051	0.4986	0.0047
Generator  Granite 4.1 8B T=1.25	5,890	0.5167	0.0010	0.0051	0.4980	0.0020

Score histograms per prompt type (panels 1–4) and generator config (panels 5–11), evaluation split. The dashed line is the threshold. Select any chart to open it at full size.

09Top-p Outliers (Llama-3.2-3B)AUROC 0.5168

DIRECTION

lower_is_ai

SWEPT FOR

fpr_0_5pct

THRESHOLD

0.0054

OVERALL AUROC

0.5168

Candidate operating points found on the validation split (the one used is highlighted):

best accuracy = 0.02804 best F1 = 0.1765 FPR ≤ 1% = 0.01002 FPR ≤ 0.5% = 0.005391 FPR ≤ 0.1% = -4.941e-324 FPR ≤ 0.01% = -4.941e-324

Subset	N	AUROC	TPR	FPR	Accuracy	F1
Overall	41,784	0.5168	0.0062	0.0071	0.4995	0.0122
Prompt  direct_reference	10,356	0.6107	0.0100	0.0077	0.5012	0.0197
Prompt  indirect_reference	10,182	0.5249	0.0128	0.0063	0.5032	0.0251
Prompt  revise	10,666	0.4953	0.0009	0.0075	0.4967	0.0019
Prompt  rewrite	10,580	0.4396	0.0013	0.0068	0.4973	0.0026
Generator  Llama 3.1 8B Instruct T=0.6 best	6,064	0.6752	0.0168	0.0053	0.5058	0.0329
Generator  Llama 3.1 8B Instruct T=1.25 worst	6,004	0.2196	0.0010	0.0083	0.4963	0.0020
Generator  Ministral 3 8B Instruct T=0.6	5,860	0.6009	0.0003	0.0075	0.4964	0.0007
Generator  Qwen3 8B T=0.7	6,118	0.6662	0.0144	0.0072	0.5036	0.0282
Generator  Gemma 4 E4B it T=0.7	5,972	0.4851	0.0040	0.0040	0.5000	0.0080
Generator  Granite 4.1 8B T=0.6	5,876	0.5588	0.0034	0.0085	0.4974	0.0067
Generator  Granite 4.1 8B T=1.25	5,890	0.4100	0.0027	0.0088	0.4969	0.0054

Score histograms per prompt type (panels 1–4) and generator config (panels 5–11), evaluation split. The dashed line is the threshold. Select any chart to open it at full size.

10Top-k Outliers (Llama-3.2-3B)AUROC 0.5168

DIRECTION

lower_is_ai

SWEPT FOR

fpr_0_5pct

THRESHOLD

0.0056

OVERALL AUROC

0.5168

Candidate operating points found on the validation split (the one used is highlighted):

best accuracy = 0.05018 best F1 = 0.8363 FPR ≤ 1% = 0.01028 FPR ≤ 0.5% = 0.005597 FPR ≤ 0.1% = -4.941e-324 FPR ≤ 0.01% = -4.941e-324

Subset	N	AUROC	TPR	FPR	Accuracy	F1
Overall	41,784	0.5168	0.0053	0.0048	0.5002	0.0104
Prompt  direct_reference	10,356	0.5622	0.0085	0.0054	0.5015	0.0168
Prompt  indirect_reference	10,182	0.5113	0.0104	0.0039	0.5032	0.0205
Prompt  revise	10,666	0.5465	0.0011	0.0054	0.4978	0.0022
Prompt  rewrite	10,580	0.4465	0.0013	0.0043	0.4985	0.0026
Generator  Llama 3.1 8B Instruct T=0.6 best	6,064	0.6778	0.0155	0.0033	0.5061	0.0304
Generator  Llama 3.1 8B Instruct T=1.25 worst	6,004	0.2099	0.0007	0.0060	0.4973	0.0013
Generator  Ministral 3 8B Instruct T=0.6	5,860	0.5863	0.0000	0.0055	0.4973	0.0000
Generator  Qwen3 8B T=0.7	6,118	0.6659	0.0147	0.0039	0.5054	0.0289
Generator  Gemma 4 E4B it T=0.7	5,972	0.4514	0.0007	0.0027	0.4990	0.0013
Generator  Granite 4.1 8B T=0.6	5,876	0.5570	0.0041	0.0071	0.4985	0.0081
Generator  Granite 4.1 8B T=1.25	5,890	0.4683	0.0007	0.0051	0.4978	0.0014

Score histograms per prompt type (panels 1–4) and generator config (panels 5–11), evaluation split. The dashed line is the threshold. Select any chart to open it at full size.

11FastDetectGPT (Llama-3.2-3B)AUROC 0.5159

DIRECTION

lower_is_ai

SWEPT FOR

fpr_0_5pct

THRESHOLD

-3.2326

OVERALL AUROC

0.5159

Candidate operating points found on the validation split (the one used is highlighted):

best accuracy = -1.812 best F1 = 8.017 FPR ≤ 1% = -2.953 FPR ≤ 0.5% = -3.233 FPR ≤ 0.1% = -4.879 FPR ≤ 0.01% = -7.346

Subset	N	AUROC	TPR	FPR	Accuracy	F1
Overall	41,784	0.5159	0.0986	0.0066	0.5460	0.1784
Prompt  direct_reference	10,356	0.4288	0.0709	0.0070	0.5320	0.1315
Prompt  indirect_reference	10,182	0.4831	0.1159	0.0075	0.5542	0.2063
Prompt  revise	10,666	0.5541	0.1001	0.0056	0.5473	0.1811
Prompt  rewrite	10,580	0.5944	0.1076	0.0062	0.5507	0.1931
Generator  Llama 3.1 8B Instruct T=0.6 worst	6,064	0.2939	0.0030	0.0082	0.4974	0.0059
Generator  Llama 3.1 8B Instruct T=1.25 best	6,004	0.8992	0.5673	0.0057	0.7808	0.7213
Generator  Ministral 3 8B Instruct T=0.6	5,860	0.3979	0.0143	0.0075	0.5034	0.0281
Generator  Qwen3 8B T=0.7	6,118	0.3076	0.0082	0.0062	0.5010	0.0161
Generator  Gemma 4 E4B it T=0.7	5,972	0.5519	0.0131	0.0064	0.5033	0.0256
Generator  Granite 4.1 8B T=0.6	5,876	0.4675	0.0099	0.0054	0.5022	0.0194
Generator  Granite 4.1 8B T=1.25	5,890	0.6924	0.0723	0.0065	0.5329	0.1341

Score histograms per prompt type (panels 1–4) and generator config (panels 5–11), evaluation split. The dashed line is the threshold. Select any chart to open it at full size.

12Entropy (Llama-3.2-3B)AUROC 0.4869

DIRECTION

lower_is_ai

SWEPT FOR

fpr_0_5pct

THRESHOLD

0.2888

OVERALL AUROC

0.4869

Candidate operating points found on the validation split (the one used is highlighted):

best accuracy = 2.151 best F1 = 8.081 FPR ≤ 1% = 0.4868 FPR ≤ 0.5% = 0.2888 FPR ≤ 0.1% = 0.208 FPR ≤ 0.01% = 0.208

Subset	N	AUROC	TPR	FPR	Accuracy	F1
Overall	41,784	0.4869	0.0005	0.0049	0.4978	0.0010
Prompt  direct_reference	10,356	0.4755	0.0004	0.0062	0.4971	0.0008
Prompt  indirect_reference	10,182	0.4783	0.0010	0.0037	0.4986	0.0020
Prompt  revise	10,666	0.5343	0.0000	0.0054	0.4973	0.0000
Prompt  rewrite	10,580	0.4567	0.0008	0.0042	0.4983	0.0015
Generator  Llama 3.1 8B Instruct T=0.6 best	6,064	0.6379	0.0003	0.0030	0.4987	0.0007
Generator  Llama 3.1 8B Instruct T=1.25 worst	6,004	0.2151	0.0003	0.0057	0.4973	0.0007
Generator  Ministral 3 8B Instruct T=0.6	5,860	0.4974	0.0000	0.0061	0.4969	0.0000
Generator  Qwen3 8B T=0.7	6,118	0.6106	0.0013	0.0042	0.4985	0.0026
Generator  Gemma 4 E4B it T=0.7	5,972	0.4135	0.0003	0.0027	0.4988	0.0007
Generator  Granite 4.1 8B T=0.6	5,876	0.5525	0.0010	0.0068	0.4971	0.0020
Generator  Granite 4.1 8B T=1.25	5,890	0.4814	0.0003	0.0058	0.4973	0.0007

Score histograms per prompt type (panels 1–4) and generator config (panels 5–11), evaluation split. The dashed line is the threshold. Select any chart to open it at full size.

13Perplexity (Llama-3.2-3B)AUROC 0.4854

DIRECTION

lower_is_ai

SWEPT FOR

fpr_0_5pct

THRESHOLD

1.2916

OVERALL AUROC

0.4854

Candidate operating points found on the validation split (the one used is highlighted):

best accuracy = 7.45 best F1 = 8377 FPR ≤ 1% = 1.54 FPR ≤ 0.5% = 1.292 FPR ≤ 0.1% = 1.139 FPR ≤ 0.01% = 1.137

Subset	N	AUROC	TPR	FPR	Accuracy	F1
Overall	41,784	0.4854	0.0005	0.0044	0.4980	0.0010
Prompt  direct_reference	10,356	0.5068	0.0004	0.0058	0.4973	0.0008
Prompt  indirect_reference	10,182	0.4852	0.0010	0.0026	0.4992	0.0020
Prompt  revise	10,666	0.5149	0.0000	0.0049	0.4976	0.0000
Prompt  rewrite	10,580	0.4338	0.0006	0.0043	0.4981	0.0011
Generator  Llama 3.1 8B Instruct T=0.6 best	6,064	0.6774	0.0007	0.0023	0.4992	0.0013
Generator  Llama 3.1 8B Instruct T=1.25 worst	6,004	0.1564	0.0003	0.0063	0.4970	0.0007
Generator  Ministral 3 8B Instruct T=0.6	5,860	0.5326	0.0000	0.0055	0.4973	0.0000
Generator  Qwen3 8B T=0.7	6,118	0.6481	0.0007	0.0029	0.4989	0.0013
Generator  Gemma 4 E4B it T=0.7	5,972	0.4061	0.0007	0.0023	0.4992	0.0013
Generator  Granite 4.1 8B T=0.6	5,876	0.5512	0.0007	0.0054	0.4976	0.0014
Generator  Granite 4.1 8B T=1.25	5,890	0.4266	0.0003	0.0061	0.4971	0.0007

Score histograms per prompt type (panels 1–4) and generator config (panels 5–11), evaluation split. The dashed line is the threshold. Select any chart to open it at full size.

■  07 — Statistics

Statistics

Univariate summaries and correlations for every statistic in the report.

Every statistic the report does arithmetic on, over the 20,892-row evaluation split. Each has N = 20,892 and no missing or non-finite values.

Distance metrics (human ↔ AI pair)

Metric	Mean	Median	Std	Min	Max
jaccard_1	0.6416	0.7434	0.2732	0.0000	1.0000
jaccard_2	0.7582	0.9006	0.2929	0.0000	1.0000
levenshtein	2308.9464	1404.0000	3256.2544	0.0000	87016.0000
softngram	0.5914	0.6754	0.3489	0.0000	1.0000
cosdist	0.2048	0.1378	0.1991	-0.0076	1.0211
bertscore	0.1455	0.1513	0.0808	-0.0000	0.4992
bertscore_precision	0.1439	0.1498	0.0825	-0.0000	0.4512
bertscore_recall	0.1459	0.1499	0.0842	-0.0000	0.5874
moverscore	0.5516	0.5975	0.2048	0.0106	1.1348
reranker	-2.8360	-4.6875	5.6218	-11.0000	18.7500

Detector scores, human vs AI text

Detector / side	Mean	Median	Std	Min	Max
EditLens Roberta-Large Score
Human	0.0565	0.0212	0.0915	0.0063	0.9995
AI	0.4735	0.4029	0.3639	0.0065	0.9996
EditLens Roberta-Large Bucket
Human	0.0605	0.0000	0.2908	0.0000	3.0000
AI	1.3514	1.0000	1.2534	0.0000	3.0000
Perplexity (Llama-3.2-3B-Instruct)
Human	18.9930	14.4788	31.4486	1.0554	2279.9307
AI	82.5218	13.0928	437.0904	1.0655	10003.4067
Perplexity (Llama-3.2-3B)
Human	13.7984	11.0648	19.7831	1.0282	1479.0899
AI	89.1387	11.1958	520.1588	1.0638	13124.9201
Entropy (Llama-3.2-3B-Instruct)
Human	2.5271	2.4894	0.6300	0.0629	7.8258
AI	2.6020	2.4018	1.1649	0.1077	9.3056
Entropy (Llama-3.2-3B)
Human	2.3825	2.4005	0.5707	0.0411	6.3306
AI	2.5570	2.4012	0.9664	0.0909	8.3789
Top-p Outliers (Llama-3.2-3B-Instruct)
Human	0.0554	0.0535	0.0163	0.0000	0.2857
AI	0.0522	0.0507	0.0173	0.0000	0.2308
Top-p Outliers (Llama-3.2-3B)
Human	0.0415	0.0414	0.0127	0.0000	0.1774
AI	0.0425	0.0401	0.0196	0.0000	0.2222
Top-k Outliers (Llama-3.2-3B-Instruct)
Human	0.1060	0.0973	0.0523	0.0000	0.6452
AI	0.1169	0.0865	0.1238	0.0000	0.8796
Top-k Outliers (Llama-3.2-3B)
Human	0.0868	0.0788	0.0474	0.0000	0.5529
AI	0.1076	0.0760	0.1235	0.0000	0.8802
FastDetectGPT (Llama-3.2-3B-Instruct)
Human	-1.6739	-1.7396	1.9618	-21.8239	25.2161
AI	-1.1973	-1.3517	1.9641	-21.0610	20.2645
FastDetectGPT (Llama-3.2-3B)
Human	-0.1766	-0.1449	1.1141	-7.3730	9.3717
AI	-0.7282	-0.2523	3.8187	-38.9644	16.2603
Binoculars
Human	0.8280	0.8516	0.1309	0.0075	1.2099
AI	0.8748	0.8787	0.0876	0.0257	1.1761

Correlation heatmap

Pearson correlation between every pair of statistics, computed over rows where both are present.

■  08 — Using the data

Using the data

Seven configs, one per generator configuration, each with a single train split.

from datasets import load_dataset, concatenate_datasets

# one config per generator configuration: shard_0 … shard_6
ds = load_dataset("G-reen/cc-2021-stat", "shard_0", split="train")

full = concatenate_datasets([
    load_dataset("G-reen/cc-2021-stat", f"shard_{i}", split="train")
    for i in range(7)
])
human, ai = full[0]["original"], full[0]["final_response"]

Columns

Column	Type	Description
Text
original	string	The human-written document.
prompt	struct	chat_turns (turn templates), use_multiturn, examples, metadata.PROMPT_TYPE.
response_0	string	Generator output for the first turn.
response_1	string	Generator output for the second turn; empty for one-turn templates.
final_response	string	The AI-side text every distance and detector score is computed on, derived from the last turn's response.
Generation
generator_model	string	Hugging Face repo of the generator.
generation_params	string (JSON)	Sampling parameters used for the row.
Distance · per pair
jaccard_1, jaccard_2, levenshtein, softngram, cosdist, bertscore, bertscore_precision, bertscore_recall, moverscore, reranker	float64	Distance or similarity between original and final_response.
Detector scores · per side
original_* · final_response_*	float64 · int64	Suffixes _perplexity_llama_{instruct,base}, _entropy_llama_*, _topp_outlier_llama_*, _topk_outlier_llama_*, _fastdetectgpt_llama_*, _binoculars, _editlens_score_roberta_large, _editlens_bucket_roberta_large (int64)

Row counts per config are under Anatomy of a pair. summary_stats.json in this repository holds machine-readable summary statistics.

■  09 — Method & provenance

Method & provenance

What you need to reproduce or interpret the numbers above.

Dataset

G-reen/cc-2021-stat · 23,214 rows in 7 configs

Human source

G-reen/cc-2021-raw: English web documents from Common Crawl CC-MAIN-2021-49

Configs

Globals config/globals.toml · analysis config/analysis_nofilter.toml

Filter conditions

None. Every row is evaluated

Split

20,892 evaluation / 2,322 validation rows (validation_size = 0.1, seed 42)

Classes

original is the human class and final_response the AI class. Each row contributes one score per side, so every detector is evaluated on 41,784 texts

Thresholds

Swept on the validation split for each detector's target (FPR ≤ 0.5% for all but EditLens Bucket, which uses best F1), then fixed for the evaluation split

Direction

Scores of lower_is_ai detectors are negated before AUROC, so every AUROC reads the same way

Subset averages

Unweighted means over the 13 detectors; the hardest subset is the one with the lowest mean TPR

Charts

Rendered from the dataset for this card. Heavy-tailed axes are clipped at the 0.5th–99.5th percentile (log-scaled for perplexity-like scores), and the clipped count is printed on each panel. Select any chart to open it at full size

Generated with FastDetector · figures and numbers from the evaluation split↑ back to top