Datasets:

G-reen
/

cc-2021-gpt-5.4-stat

Modalities:

Formats:

parquet

Size:

Libraries:

Dataset card Data Studio Files Files and versions

xet

Community

Dataset Viewer

Auto-converted to Parquet Duplicate

Split (1)

train · 500 rows

original string	prompt dict	response_0 string	response_1 string	final_response string	generator_model string	generation_params string	jaccard_1 float64	jaccard_2 float64	levenshtein float64	cosdist float64	original_editlens_score_cc2020_best float64	final_response_editlens_bucket_cc2020_best int64	final_response_editlens_score_cc2020_best float64	original_editlens_score_cc2020_2bin float64	final_response_editlens_bucket_cc2020_2bin int64	final_response_editlens_score_cc2020_2bin float64	original_editlens_score_cosdist_2bin float64	final_response_editlens_bucket_cosdist_2bin int64	final_response_editlens_score_cosdist_2bin float64	original_editlens_score_cosdist_4bin float64	final_response_editlens_bucket_cosdist_4bin int64	final_response_editlens_score_cosdist_4bin float64	original_editlens_score_cc2020_graded float64	final_response_editlens_bucket_cc2020_graded int64	final_response_editlens_score_cc2020_graded float64	original_editlens_score_cosdist_graded float64	final_response_editlens_bucket_cosdist_graded int64	final_response_editlens_score_cosdist_graded float64
Deciding on what to do with the budget you've set aside to renovate your bathroom can be tough. There are improvements you want to improve your lifestyle and there are improvements that could improve your home's resale value. Ideally, you should choose improvements that fit both criteria. Since personal preference vari...	{ "chat_turns": [ "{{DOC}}\nChange the tone of this text to be more relaxed and friendly.\nOutput the full new text with no extra statements or commentations." ], "use_multiturn": false, "examples": [], "metadata": { "PROMPT_TYPE": "revise" } }	Figuring out how to spend the budget you’ve set aside for a bathroom renovation can feel a little overwhelming. There are upgrades that can make your everyday routine more enjoyable, and others that might help boost your home’s resale value. The sweet spot is choosing updates that do both. Since everyone’s style and pr...		Figuring out how to spend the budget you’ve set aside for a bathroom renovation can feel a little overwhelming. There are upgrades that can make your everyday routine more enjoyable, and others that might help boost your home’s resale value. The sweet spot is choosing updates that do both. Since everyone’s style and pr...	gpt-5.4	{"disable_thinking": true, "is_api_model": true}	0.695652	0.898734	287	0.083218	0.093558	0	0.018356	0.183009	0	0.093347	0.2763	0	0.184473	0.185617	0	0.126078	0.315728	0	0.129965	0.150131	0	0.073344
I am warm, caring and straightforward, with a deep respect for my client’s struggles and strengths. I value my client’s perspectives and insights and together we try to come to deeper levels of understanding. Individuals face their issues at their own pace, as they explore the roots of their present suffering. As thera...	{ "chat_turns": [ "{{DOC}}\nModify this text to be as long as possible. However, the new text must contain at least half of the original text, unmodified, repeated verbatim.\nOutput the full new text with no extra statements or commentations." ], "use_multiturn": false, "examples": [], "metadata": { "...	I am warm, caring and straightforward, with a deep respect for my client’s struggles and strengths. I value my client’s perspectives and insights and together we try to come to deeper levels of understanding. Individuals face their issues at their own pace, as they explore the roots of their present suffering. As thera...		I am warm, caring and straightforward, with a deep respect for my client’s struggles and strengths. I value my client’s perspectives and insights and together we try to come to deeper levels of understanding. Individuals face their issues at their own pace, as they explore the roots of their present suffering. As thera...	gpt-5.4	{"disable_thinking": true, "is_api_model": true}	0.742033	0.807815	9,796	0.065283	0.009749	0	0.021994	0.066085	0	0.166134	0.060087	0	0.072637	0.034811	0	0.048048	0.003589	0	0.009278	0.016039	0	0.230625
2016 is looking like a good year for off the wall golf games. We've had the announcement of 100ft Robot Golf where you hit the links as a giant mecha and can flatten cities to clear a path to the green and now we have Dangerous Golf, where the goal…	{ "chat_turns": [ "{{DOC}}\nTake 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.\nOutput the full new text with no extra statements or commentation...	2026 is shaping up to be a weirdly brilliant year for racing games. Weve already seen the reveal of Turbo Island, where you drift tiny hatchbacks around supermarket aisles at 70mph while knocking over cereal displays for boost, and now theres Night Shift Taxi, a game that seems to ask the very important question: what ...		2026 is shaping up to be a weirdly brilliant year for racing games. Weve already seen the reveal of Turbo Island, where you drift tiny hatchbacks around supermarket aisles at 70mph while knocking over cereal displays for boost, and now theres Night Shift Taxi, a game that seems to ask the very important question: what ...	gpt-5.4	{"disable_thinking": true, "is_api_model": true}	0.948097	0.990268	1,931	0.568605	0.044884	0	0.035654	0.112795	0	0.281011	0.18462	0	0.154695	0.041807	0	0.089862	0.061212	6	0.876569	0.086083	10	0.763511
The ICO reiterates the EDPB's FAQs which state a risk assessment is required when relying on the Standard Contractual Clauses as to whether they provide sufficient protection within the legal framework of the third country the data is being transferred to, whether it is to the US or elsewhere. The EDPB confirmed in the...	{ "chat_turns": [ "{{DOC}}\nChange the tone of this text to be more empathetic and understanding.\nOutput the full new text with no extra statements or commentations." ], "use_multiturn": false, "examples": [], "metadata": { "PROMPT_TYPE": "revise" } }	The ICO recognises the EDPB’s FAQs, which explain that organisations relying on the Standard Contractual Clauses must carry out a risk assessment to determine whether those clauses provide sufficient protection within the legal framework of the third country to which personal data is being transferred, whether that is ...		The ICO recognises the EDPB’s FAQs, which explain that organisations relying on the Standard Contractual Clauses must carry out a risk assessment to determine whether those clauses provide sufficient protection within the legal framework of the third country to which personal data is being transferred, whether that is ...	gpt-5.4	{"disable_thinking": true, "is_api_model": true}	0.438889	0.601881	534	0.023947	0.052677	0	0.033466	0.203075	1	0.732974	0.232659	0	0.263173	0.116161	0	0.139971	0.138519	0	0.1639	0.061156	0	0.238906
Vv Magazine’s Associate Editor Azra Hirji will be the first to tell you that, as a woman of colour working in a predominantly ‘white’ industry where having strong beauty game garners you some serious brownie points, it’s often difficult to find a little beauty inspiration that suits her skin tone. Azra decided to take ...	{ "chat_turns": [ "{{DOC}}\nWrite a detailed descriptor for the style of writing of this text, including its qualitative and quantitative properties.\nDo not output anything besides what you were requested to write, and do not output any extra commentary.", "{{RESP_0}}\nAbove is an AI generated descriptor/tra...	The writing style is editorial-feature journalism with a strong lifestyle/beauty-magazine orientation, combining a conversational introduction with a structured interview format. It is designed for broad, style-conscious readers and balances personal narrative, aspirational beauty culture, and practical consumer advice...	As a woman of colour, finding beauty inspiration and advice that actually speaks to your skin tone and your experiences can be frustrating in a predominantly “white” industry. That’s one of the reasons I started this series: to spotlight women from different backgrounds, with gorgeous features and real beauty stories, ...	As a woman of colour, finding beauty inspiration and advice that actually speaks to your skin tone and your experiences can be frustrating in a predominantly “white” industry. That’s one of the reasons I started this series: to spotlight women from different backgrounds, with gorgeous features and real beauty stories, ...	gpt-5.4	{"disable_thinking": true, "is_api_model": true}	0.778017	0.936803	2,195	0.194041	0.005308	2	0.669371	0.054601	1	0.998716	0.045863	1	0.999957	0.032941	2	0.657726	0.006827	1	0.263344	0.011959	3	0.353859
We are available anywhere and anytime and we are one of the leading carpentry companies on the market right now. No matter how big or how small the jobs are, from Industrial to Residential, we assist all customers; there's nothing we can't do. Currently we have many well trained, professional and reliable carpentry bus...	{ "chat_turns": [ "{{DOC}}\nModify this text to be as long as possible. However, the new text must contain at least half of the original text, unmodified, repeated verbatim.\nOutput the full new text with no extra statements or commentations.", "{{RESP_0}}\nModify this text to be as long as possible. However,...	We are available anywhere and anytime and we are one of the leading carpentry companies on the market right now. No matter how big or how small the jobs are, from Industrial to Residential, we assist all customers; there's nothing we can't do. Currently we have many well trained, professional and reliable carpentry bus...	We are available anywhere and anytime and we are one of the leading carpentry companies on the market right now. No matter how big or how small the jobs are, from Industrial to Residential, we assist all customers; there's nothing we can't do. Currently we have many well trained, professional and reliable carpentry bus...	We are available anywhere and anytime and we are one of the leading carpentry companies on the market right now. No matter how big or how small the jobs are, from Industrial to Residential, we assist all customers; there's nothing we can't do. Currently we have many well trained, professional and reliable carpentry bus...	gpt-5.4	{"disable_thinking": true, "is_api_model": true}	0.761578	0.845768	18,068	0.061773	0.013157	2	0.368791	0.094678	1	0.958226	0.153803	1	0.993307	0.084614	2	0.647201	0.015182	3	0.410649	0.059305	3	0.296228
Vintage 1993 Tiger FULL HOUSE Dolls Danny's FAMILY. It is new in sealed package. There is some shelf wear and some sticker residue on the box. Otherwise the box is in nice shape. Please see photos for condition and what you will be getting. Items from a smoke free home.	{ "chat_turns": [ "{{DOC}}\nEdit this text by replacing, removing or inserting twenty to thirty words, without changing anything else.\nOutput the full new text with no extra statements or commentations.", "{{RESP_0}}\nEdit this text by replacing, removing or inserting twenty to thirty words, without changing...		Please provide the text you want edited.	Please provide the text you want edited.	gpt-5.4	{"disable_thinking": true, "is_api_model": true}	0.9375	1	238	0.714301	0.022975	2	0.399077	0.058346	0	0.261708	0.143908	1	0.696784	0.054185	2	0.491295	0.094705	0	0.542752	0.045643	0	0.498472
"agen bola betting occurs illegally as well as legally on this game in any way levels of perform. Wh(...TRUNCATED)	{"chat_turns":["{{DOC}}\nModify this text to be as long as possible. However, the new text must cont(...TRUNCATED)	"agen bola betting occurs illegally as well as legally on this game in any way levels of perform. Wh(...TRUNCATED)		"agen bola betting occurs illegally as well as legally on this game in any way levels of perform. Wh(...TRUNCATED)	gpt-5.4	{"disable_thinking": true, "is_api_model": true}	0.006228	0.005843	34,713	0.076881	0.006038	0	0.006038	0.048858	0	0.048858	0.07836	0	0.07836	0.047649	0	0.047649	0.00666	0	0.00666	0.020061	0	0.020061
"Join us to pray live via Zoom videoconference at 8:00 a.m. ET using the following link: Zoom Link f(...TRUNCATED)	{"chat_turns":["{{DOC}}\nWrite a new piece of text that is on a completely different topic, but reus(...TRUNCATED)	"Join us to plan live via Zoom videoconference at 8:00 a.m. ET using the following link: Zoom Link f(...TRUNCATED)		"Join us to plan live via Zoom videoconference at 8:00 a.m. ET using the following link: Zoom Link f(...TRUNCATED)	gpt-5.4	{"disable_thinking": true, "is_api_model": true}	0.523288	0.615181	5,683	0.42339	0.016927	0	0.099364	0.060087	0	0.38122	0.067056	1	0.882428	0.045607	2	0.530983	0.03253	0	0.109598	0.047569	0	0.417338
"The U.S. Congress created the EB-5 Program in 1990 in order to enable high net worth foreign invest(...TRUNCATED)	{"chat_turns":["{{DOC}}\nTranslate this text to Chinese.\nDo not output anything besides what you we(...TRUNCATED)	"美国国会于1990年设立EB-5项目，旨在使高净值外国投资者能够通过以有利于(...TRUNCATED)	"The U.S. Congress established the EB-5 Program in 1990 to enable high-net-worth foreign investors t(...TRUNCATED)	"The U.S. Congress established the EB-5 Program in 1990 to enable high-net-worth foreign investors t(...TRUNCATED)	gpt-5.4	{"disable_thinking": true, "is_api_model": true}	0.564246	0.764468	2,551	0.085854	0.006376	0	0.011625	0.013223	0	0.091382	0.011332	0	0.072112	0.006474	0	0.04647	0.00298	0	0.016735	0.00384	0	0.047331

End of preview. Expand in Data Studio

Auto-Generated FastDetector Dataset

Dataset: G-reen/cc-2021-gpt-5.4-stat
Globals Config: config/globals_cc2021_gpt54.toml
Analysis Config: config/analysis_gpt54.toml
Rows Loaded: 500
Filter Conditions: cosdist >= 0.03
Rows Analyzed (after filtering): 375
Evaluation / Validation Rows: 337 / 38 (validation_size = 0.1)
Base Columns: original (Human), final_response (AI)
Distance Metrics: jaccard_1, jaccard_2, levenshtein, cosdist
Distance Metrics Skipped (not in this dataset): jaccard_3, softngram, bertscore, bertscore_precision, bertscore_recall, moverscore, reranker
Prompt Subsets: 4 (direct_reference, indirect_reference, revise, rewrite)
Generator Configs: 1 (gpt-5.4 (Temp: Unknown))
Classifiers:
- ~~EditLens Roberta-Large Score~~ - SKIPPED, the dataset has no original_editlens_score_roberta_large, final_response_editlens_score_roberta_large column(s)
- ~~EditLens Roberta-Large Bucket~~ - SKIPPED, the dataset has no original_editlens_bucket_roberta_large, final_response_editlens_bucket_roberta_large column(s)
- ~~Perplexity (Llama-3.2-3B-Instruct)~~ - SKIPPED, the dataset has no original_perplexity_llama_instruct, final_response_perplexity_llama_instruct column(s)
- ~~Perplexity (Llama-3.2-3B)~~ - SKIPPED, the dataset has no original_perplexity_llama_base, final_response_perplexity_llama_base column(s)
- ~~Entropy (Llama-3.2-3B-Instruct)~~ - SKIPPED, the dataset has no original_entropy_llama_instruct, final_response_entropy_llama_instruct column(s)
- ~~Entropy (Llama-3.2-3B)~~ - SKIPPED, the dataset has no original_entropy_llama_base, final_response_entropy_llama_base column(s)
- ~~Top-p Outliers (Llama-3.2-3B-Instruct)~~ - SKIPPED, the dataset has no original_topp_outlier_llama_instruct, final_response_topp_outlier_llama_instruct column(s)
- ~~Top-p Outliers (Llama-3.2-3B)~~ - SKIPPED, the dataset has no original_topp_outlier_llama_base, final_response_topp_outlier_llama_base column(s)
- ~~Top-k Outliers (Llama-3.2-3B-Instruct)~~ - SKIPPED, the dataset has no original_topk_outlier_llama_instruct, final_response_topk_outlier_llama_instruct column(s)
- ~~Top-k Outliers (Llama-3.2-3B)~~ - SKIPPED, the dataset has no original_topk_outlier_llama_base, final_response_topk_outlier_llama_base column(s)
- ~~FastDetectGPT (Llama-3.2-3B-Instruct)~~ - SKIPPED, the dataset has no original_fastdetectgpt_llama_instruct, final_response_fastdetectgpt_llama_instruct column(s)
- ~~FastDetectGPT (Llama-3.2-3B)~~ - SKIPPED, the dataset has no original_fastdetectgpt_llama_base, final_response_fastdetectgpt_llama_base column(s)
- ~~Binoculars~~ - SKIPPED, the dataset has no original_binoculars, final_response_binoculars column(s)
- EditLens Llama-3.2-3B cc-2020 2-Bucket Score - columns *_editlens_score_cc2020_2bin, direction higher_is_ai (swept for fpr_0_5pct on the validation split)
- EditLens Llama-3.2-3B cc-2020 2-Bucket Bucket - columns *_editlens_bucket_cc2020_2bin, direction higher_is_ai (swept for f1 on the validation split)
- EditLens Llama-3.2-3B cosdist 2-Bucket Score - columns *_editlens_score_cosdist_2bin, direction higher_is_ai (swept for fpr_0_5pct on the validation split)
- EditLens Llama-3.2-3B cosdist 2-Bucket Bucket - columns *_editlens_bucket_cosdist_2bin, direction higher_is_ai (swept for f1 on the validation split)
- EditLens Llama-3.2-3B cc-2020 Graded (7 buckets) Score - columns *_editlens_score_cc2020_graded, direction higher_is_ai (swept for fpr_0_5pct on the validation split)
- EditLens Llama-3.2-3B cc-2020 Graded (7 buckets) Bucket - columns *_editlens_bucket_cc2020_graded, direction higher_is_ai (swept for f1 on the validation split)
- EditLens Llama-3.2-3B cosdist Graded (12 buckets) Score - columns *_editlens_score_cosdist_graded, direction higher_is_ai (swept for fpr_0_5pct on the validation split)
- EditLens Llama-3.2-3B cosdist Graded (12 buckets) Bucket - columns *_editlens_bucket_cosdist_graded, direction higher_is_ai (swept for f1 on the validation split)

This readme computes the detection report for G-reen/cc-2021-gpt-5.4-stat: 375 human/AI text pairs from 1 generator configuration(s) and 4 prompt type(s), summarised univariately, correlated against each other, and used to score 8 classifier(s) as AI-text detectors. Over the whole evaluation split EditLens Llama-3.2-3B cosdist Graded (12 buckets) Score separates the two classes best (AUROC 0.8696, catching 56.08% of AI rows at 3.56% false positives), while EditLens Llama-3.2-3B cc-2020 Graded (7 buckets) Bucket is weakest (AUROC 0.6581). 4 pairwise distance metric(s) (jaccard_1, jaccard_2, levenshtein, cosdist) measure how far each AI response moved from its human original; they are profiled here and correlated against every classifier score. Every classifier is then broken down over 4 prompt subset(s). For EditLens Llama-3.2-3B cc-2020 2-Bucket Score, the easiest subset is revise (AUROC 0.8689) and the hardest direct_reference (AUROC 0.7371). The final section reports EditLens Llama-3.2-3B cc-2020 2-Bucket Score across the 1 generator configuration(s) that wrote the AI side of the corpus.

Univariate Analysis
Correlation Heatmap
Histogram, Distances
- Distance Histograms per Prompt Subset
Histogram, Classification
Classifiers Comparison Table
Classifier Thresholds
Classifier Report: EditLens Llama-3.2-3B cc-2020 2-Bucket Score
- EditLens Llama-3.2-3B cc-2020 2-Bucket Score: By Prompt Subset
- EditLens Llama-3.2-3B cc-2020 2-Bucket Score: Score Histograms per Prompt Subset
Classifier Report: EditLens Llama-3.2-3B cc-2020 2-Bucket Bucket
- EditLens Llama-3.2-3B cc-2020 2-Bucket Bucket: By Prompt Subset
- EditLens Llama-3.2-3B cc-2020 2-Bucket Bucket: Score Histograms per Prompt Subset
Classifier Report: EditLens Llama-3.2-3B cosdist 2-Bucket Score
- EditLens Llama-3.2-3B cosdist 2-Bucket Score: By Prompt Subset
- EditLens Llama-3.2-3B cosdist 2-Bucket Score: Score Histograms per Prompt Subset
Classifier Report: EditLens Llama-3.2-3B cosdist 2-Bucket Bucket
- EditLens Llama-3.2-3B cosdist 2-Bucket Bucket: By Prompt Subset
- EditLens Llama-3.2-3B cosdist 2-Bucket Bucket: Score Histograms per Prompt Subset
Classifier Report: EditLens Llama-3.2-3B cc-2020 Graded (7 buckets) Score
- EditLens Llama-3.2-3B cc-2020 Graded (7 buckets) Score: By Prompt Subset
- EditLens Llama-3.2-3B cc-2020 Graded (7 buckets) Score: Score Histograms per Prompt Subset
Classifier Report: EditLens Llama-3.2-3B cc-2020 Graded (7 buckets) Bucket
- EditLens Llama-3.2-3B cc-2020 Graded (7 buckets) Bucket: By Prompt Subset
- EditLens Llama-3.2-3B cc-2020 Graded (7 buckets) Bucket: Score Histograms per Prompt Subset
Classifier Report: EditLens Llama-3.2-3B cosdist Graded (12 buckets) Score
- EditLens Llama-3.2-3B cosdist Graded (12 buckets) Score: By Prompt Subset
- EditLens Llama-3.2-3B cosdist Graded (12 buckets) Score: Score Histograms per Prompt Subset
Classifier Report: EditLens Llama-3.2-3B cosdist Graded (12 buckets) Bucket
- EditLens Llama-3.2-3B cosdist Graded (12 buckets) Bucket: By Prompt Subset
- EditLens Llama-3.2-3B cosdist Graded (12 buckets) Bucket: Score Histograms per Prompt Subset
Manually Specified Full Report
- Score Histograms per Generator Config
- Distance Histograms per Generator Config

Univariate Analysis

Every statistic the report does arithmetic on, over the 337-row evaluation split. Invalid/Error counts rows whose value is missing or non-finite; those rows are excluded from the other columns.

Statistic	N	Mean	Median	Std Dev	Min	Max
jaccard_1	337	0.6152	0.6898	0.2345	0.0000	1.0000
jaccard_2	337	0.7519	0.8530	0.2518	0.0011	1.0000
levenshtein	337	3186.3739	1629.0000	4104.4044	115.0000	34713.0000
cosdist	337	0.1973	0.1200	0.1846	0.0303	0.9600
EditLens Llama-3.2-3B cc-2020 2-Bucket Score (Human)	337	0.1444	0.0711	0.1819	0.0103	0.9588
EditLens Llama-3.2-3B cc-2020 2-Bucket Score (AI)	337	0.4782	0.4033	0.3529	0.0213	0.9999
EditLens Llama-3.2-3B cc-2020 2-Bucket Bucket (Human)	337	0.0623	0.0000	0.2417	0.0000	1.0000
EditLens Llama-3.2-3B cc-2020 2-Bucket Bucket (AI)	337	0.4599	0.0000	0.4984	0.0000	1.0000
EditLens Llama-3.2-3B cosdist 2-Bucket Score (Human)	337	0.1188	0.0781	0.1226	0.0208	0.9121
EditLens Llama-3.2-3B cosdist 2-Bucket Score (AI)	337	0.5188	0.4545	0.3686	0.0206	1.0000
EditLens Llama-3.2-3B cosdist 2-Bucket Bucket (Human)	337	0.0326	0.0000	0.1777	0.0000	1.0000
EditLens Llama-3.2-3B cosdist 2-Bucket Bucket (AI)	337	0.4748	0.0000	0.4994	0.0000	1.0000
EditLens Llama-3.2-3B cc-2020 Graded (7 buckets) Score (Human)	337	0.0337	0.0103	0.0661	0.0019	0.5048
EditLens Llama-3.2-3B cc-2020 Graded (7 buckets) Score (AI)	337	0.2405	0.0753	0.2809	0.0022	0.9686
EditLens Llama-3.2-3B cc-2020 Graded (7 buckets) Bucket (Human)	337	0.0356	0.0000	0.3249	0.0000	3.0000
EditLens Llama-3.2-3B cc-2020 Graded (7 buckets) Bucket (AI)	337	1.2463	0.0000	1.9493	0.0000	6.0000
EditLens Llama-3.2-3B cosdist Graded (12 buckets) Score (Human)	337	0.0366	0.0229	0.0423	0.0057	0.3018
EditLens Llama-3.2-3B cosdist Graded (12 buckets) Score (AI)	337	0.2560	0.1926	0.2224	0.0073	0.9294
EditLens Llama-3.2-3B cosdist Graded (12 buckets) Bucket (Human)	337	0.0059	0.0000	0.1088	0.0000	2.0000
EditLens Llama-3.2-3B cosdist Graded (12 buckets) Bucket (AI)	337	2.1840	0.0000	3.1257	0.0000	11.0000

Correlation Heatmap

Pearson correlation between every statistic of interest, computed over the rows where both statistics are present.

Histogram, Distances

Distribution of each pairwise distance between a human original and its AI rewrite, over the whole evaluation split.

Distance Histograms per Prompt Subset

The same distances, split by the 4 prompt subset(s).

Histogram, Classification

Human and AI score distributions for each classifier, overlaid, over the whole evaluation split.

Classifiers Comparison Table

AUROC is threshold-free; TPR, FPR, accuracy and F1 are measured at each classifier's own pinned threshold (shown in the first column). Rows a classifier produced no usable score for are excluded from its metrics and counted in the univariate table's Invalid/Error column.

Classifier	Threshold	AUROC	TPR @ Threshold	FPR @ Threshold	Accuracy	F1
EditLens Llama-3.2-3B cc-2020 2-Bucket Score	0.9300	0.8032	0.2018	0.0030	0.5994	0.3350
EditLens Llama-3.2-3B cc-2020 2-Bucket Bucket	0.0000	0.6988	0.4599	0.0623	0.6988	0.6043
EditLens Llama-3.2-3B cosdist 2-Bucket Score	0.7702	0.8383	0.3680	0.0030	0.6825	0.5368
EditLens Llama-3.2-3B cosdist 2-Bucket Bucket	0.0000	0.7211	0.4748	0.0326	0.7211	0.6299
EditLens Llama-3.2-3B cc-2020 Graded (7 buckets) Score	0.4874	0.7739	0.2226	0.0030	0.6098	0.3632
❗ EditLens Llama-3.2-3B cc-2020 Graded (7 buckets) Bucket	0.0000	0.6581	0.3264	0.0119	0.6573	0.4878
✔️ EditLens Llama-3.2-3B cosdist Graded (12 buckets) Score	0.1576	0.8696	0.5608	0.0356	0.7626	0.7026
EditLens Llama-3.2-3B cosdist Graded (12 buckets) Bucket	0.0000	0.7202	0.4421	0.0030	0.7196	0.6119

✔️ marks the best AUROC, ❗ the worst.

Classifier Thresholds

Accuracy against threshold on the 38-row validation split, with each candidate threshold type marked. The type named in the run configuration is the one pinned for the tables above.

EditLens Llama-3.2-3B cc-2020 2-Bucket Score (swept for fpr_0_5pct on the validation split)

EditLens Llama-3.2-3B cc-2020 2-Bucket Bucket (swept for f1 on the validation split)

EditLens Llama-3.2-3B cosdist 2-Bucket Score (swept for fpr_0_5pct on the validation split)

EditLens Llama-3.2-3B cosdist 2-Bucket Bucket (swept for f1 on the validation split)

EditLens Llama-3.2-3B cc-2020 Graded (7 buckets) Score (swept for fpr_0_5pct on the validation split)

EditLens Llama-3.2-3B cc-2020 Graded (7 buckets) Bucket (swept for f1 on the validation split)

EditLens Llama-3.2-3B cosdist Graded (12 buckets) Score (swept for fpr_0_5pct on the validation split)

EditLens Llama-3.2-3B cosdist Graded (12 buckets) Bucket (swept for f1 on the validation split)

Classifier Report: EditLens Llama-3.2-3B cc-2020 2-Bucket Score

Threshold swept for fpr_0_5pct on the validation split = 0.9300; scores read from *_editlens_score_cc2020_2bin (higher_is_ai).

EditLens Llama-3.2-3B cc-2020 2-Bucket Score: By Prompt Subset

Prompt Subset	N	AUROC	TPR	FPR	Accuracy	F1
Overall	674	0.8032	0.2018	0.0030	0.5994	0.3350
❗ direct_reference	192	0.7371	0.1667	0.0104	0.5781	0.2832
indirect_reference	208	0.8200	0.1635	0.0000	0.5817	0.2810
✔️ revise	158	0.8689	0.3418	0.0000	0.6709	0.5094
rewrite	116	0.8230	0.1379	0.0000	0.5690	0.2424

EditLens Llama-3.2-3B cc-2020 2-Bucket Score: Score Histograms per Prompt Subset

Classifier Report: EditLens Llama-3.2-3B cc-2020 2-Bucket Bucket

Threshold swept for f1 on the validation split = 0.0000; scores read from *_editlens_bucket_cc2020_2bin (higher_is_ai).

EditLens Llama-3.2-3B cc-2020 2-Bucket Bucket: By Prompt Subset

Prompt Subset	N	AUROC	TPR	FPR	Accuracy	F1
Overall	674	0.6988	0.4599	0.0623	0.6988	0.6043
❗ direct_reference	192	0.6615	0.4375	0.1146	0.6615	0.5638
indirect_reference	208	0.6923	0.4135	0.0288	0.6923	0.5733
✔️ revise	158	0.7722	0.6076	0.0633	0.7722	0.7273
rewrite	116	0.6724	0.3793	0.0345	0.6724	0.5366

EditLens Llama-3.2-3B cc-2020 2-Bucket Bucket: Score Histograms per Prompt Subset

Classifier Report: EditLens Llama-3.2-3B cosdist 2-Bucket Score

Threshold swept for fpr_0_5pct on the validation split = 0.7702; scores read from *_editlens_score_cosdist_2bin (higher_is_ai).

EditLens Llama-3.2-3B cosdist 2-Bucket Score: By Prompt Subset

Prompt Subset	N	AUROC	TPR	FPR	Accuracy	F1
Overall	674	0.8383	0.3680	0.0030	0.6825	0.5368
❗ direct_reference	192	0.7764	0.3021	0.0000	0.6510	0.4640
indirect_reference	208	0.8629	0.3269	0.0000	0.6635	0.4928
✔️ revise	158	0.9120	0.5696	0.0127	0.7785	0.7200
rewrite	116	0.8081	0.2759	0.0000	0.6379	0.4324

EditLens Llama-3.2-3B cosdist 2-Bucket Score: Score Histograms per Prompt Subset

Classifier Report: EditLens Llama-3.2-3B cosdist 2-Bucket Bucket

Threshold swept for f1 on the validation split = 0.0000; scores read from *_editlens_bucket_cosdist_2bin (higher_is_ai).

EditLens Llama-3.2-3B cosdist 2-Bucket Bucket: By Prompt Subset

Prompt Subset	N	AUROC	TPR	FPR	Accuracy	F1
Overall	674	0.7211	0.4748	0.0326	0.7211	0.6299
❗ direct_reference	192	0.6667	0.3646	0.0312	0.6667	0.5224
indirect_reference	208	0.7019	0.4423	0.0385	0.7019	0.5974
✔️ revise	158	0.8101	0.6709	0.0506	0.8101	0.7794
rewrite	116	0.7241	0.4483	0.0000	0.7241	0.6190

EditLens Llama-3.2-3B cosdist 2-Bucket Bucket: Score Histograms per Prompt Subset

Classifier Report: EditLens Llama-3.2-3B cc-2020 Graded (7 buckets) Score

Threshold swept for fpr_0_5pct on the validation split = 0.4874; scores read from *_editlens_score_cc2020_graded (higher_is_ai).

EditLens Llama-3.2-3B cc-2020 Graded (7 buckets) Score: By Prompt Subset

Prompt Subset	N	AUROC	TPR	FPR	Accuracy	F1
Overall	674	0.7739	0.2226	0.0030	0.6098	0.3632
❗ direct_reference	192	0.6949	0.2396	0.0104	0.6146	0.3833
indirect_reference	208	0.8142	0.1250	0.0000	0.5625	0.2222
✔️ revise	158	0.8553	0.3924	0.0000	0.6962	0.5636
rewrite	116	0.7323	0.1379	0.0000	0.5690	0.2424

EditLens Llama-3.2-3B cc-2020 Graded (7 buckets) Score: Score Histograms per Prompt Subset

Classifier Report: EditLens Llama-3.2-3B cc-2020 Graded (7 buckets) Bucket

Threshold swept for f1 on the validation split = 0.0000; scores read from *_editlens_bucket_cc2020_graded (higher_is_ai).

EditLens Llama-3.2-3B cc-2020 Graded (7 buckets) Bucket: By Prompt Subset

Prompt Subset	N	AUROC	TPR	FPR	Accuracy	F1
Overall	674	0.6581	0.3264	0.0119	0.6573	0.4878
direct_reference	192	0.6429	0.3021	0.0208	0.6406	0.4567
indirect_reference	208	0.6208	0.2596	0.0192	0.6202	0.4060
✔️ revise	158	0.7532	0.5063	0.0000	0.7532	0.6723
❗ rewrite	116	0.6207	0.2414	0.0000	0.6207	0.3889

EditLens Llama-3.2-3B cc-2020 Graded (7 buckets) Bucket: Score Histograms per Prompt Subset

Classifier Report: EditLens Llama-3.2-3B cosdist Graded (12 buckets) Score

Threshold swept for fpr_0_5pct on the validation split = 0.1576; scores read from *_editlens_score_cosdist_graded (higher_is_ai).

EditLens Llama-3.2-3B cosdist Graded (12 buckets) Score: By Prompt Subset

Prompt Subset	N	AUROC	TPR	FPR	Accuracy	F1
Overall	674	0.8696	0.5608	0.0356	0.7626	0.7026
direct_reference	192	0.8802	0.5625	0.0312	0.7656	0.7059
indirect_reference	208	0.8871	0.5288	0.0288	0.7500	0.6790
✔️ revise	158	0.9039	0.6962	0.0633	0.8165	0.7914
❗ rewrite	116	0.7748	0.4310	0.0172	0.7069	0.5952

EditLens Llama-3.2-3B cosdist Graded (12 buckets) Score: Score Histograms per Prompt Subset

Classifier Report: EditLens Llama-3.2-3B cosdist Graded (12 buckets) Bucket

Threshold swept for f1 on the validation split = 0.0000; scores read from *_editlens_bucket_cosdist_graded (higher_is_ai).

EditLens Llama-3.2-3B cosdist Graded (12 buckets) Bucket: By Prompt Subset

Prompt Subset	N	AUROC	TPR	FPR	Accuracy	F1
Overall	674	0.7202	0.4421	0.0030	0.7196	0.6119
direct_reference	192	0.7083	0.4167	0.0000	0.7083	0.5882
indirect_reference	208	0.7019	0.4038	0.0000	0.7019	0.5753
✔️ revise	158	0.8137	0.6329	0.0127	0.8101	0.7692
❗ rewrite	116	0.6466	0.2931	0.0000	0.6466	0.4533

EditLens Llama-3.2-3B cosdist Graded (12 buckets) Bucket: Score Histograms per Prompt Subset

Manually Specified Full Report

This section is hardcoded: it reports the first configured classifier (EditLens Llama-3.2-3B cc-2020 2-Bucket Score) over the generator model/sampling configurations that produced the AI side of each pair.

Generator Config	N	AUROC	TPR	FPR	Accuracy	F1
Overall	674	0.8032	0.2018	0.0030	0.5994	0.3350
✔️ gpt-5.4 (Temp: Unknown)	674	0.8032	0.2018	0.0030	0.5994	0.3350

Score Histograms per Generator Config

Distance Histograms per Generator Config

Downloads last month: 88