Spaces:
Running
Running
Commit History
2 days ago
-
Stronger AI detector: two more paraphrase-robust models averaged into the score
unknown +1: Claude Opus 5.5Nine public detectors were tested (devtools/compare_detectors.py) on 18 AI-written academic samples, 21 human abstracts from 2008 and 16 paragraphs our humanizer had rewritten. Added: ShantanuT01/vanguard-ai-text-detector (ModernBERT-large, MIT) and Oxidane/tmr-ai-text-detector (RoBERTa, MIT). The final score is the mean of the previous ensemble and the two: AI 18/18, human 0/21, humanized 11/16 flagged (was 6/16). Rejected: PirateXX (flagged every human text), AICodexLab ModernBERT (18/21 human), desklib academic (13/21), econ-distilroberta (7/21), vraj33 (flagged nothing). DETECTOR_EXTRA sets the extra models ("" = none). Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>dc37bdf -
Fix start-up: the half-precision check is started at the end of detector.py (it was called before its definition)
unknown +1: Claude Opus 5.5Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>306f02a -
One-minute deadline for the fast engine; half-precision detector on the GPU when it gives the same verdicts
unknown +1: Claude Opus 5.5- fast.py: the whole job ends within FAST_BUDGET (52 s) whatever the length: language-model requests still running at the deadline are dropped (their sentences keep their best version), rounds are planned from measured times, and long documents' candidates are pre-ranked on the GPU too. - detector.py: half precision on the GPU (about 2x faster), switched on at start-up only if it changes no verdict and no score by more than 0.05 on the bundled samples; logged either way. DETECTOR_HALF=0 = never. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>439ec70
3 days ago
-
Fast engine on CPU about a third faster: versions pre-ranked by the two small detector models
unknown +1: Claude Opus 5.5On a CPU the large DeBERTa detector was about 2/3 of the detector time. Every version is now ranked by the two small models (on its own), the fluency check runs on the shortlist, and only the 2 best per sentence go to the full three-model detector; the document's AI score is still always the full detector's. Benchmark (4 AI paragraphs, laptop CPU): 100% -> 0% in 3 rounds in 218 s (was 320-350 s). int8 quantisation of the detector was tested and rejected: it changed 8 of 77 verdicts (devtools/bench_int8_detector.py). Capping versions per model at 2 was tested and rejected: it stalled at 23%. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>bc66a36 -
Stronger AI detector: paraphrase-robust third model in the ensemble
unknown +1: Claude Opus 5.5fakespot-ai/roberta-base-ai-text-detection-v1 averaged with the current coai + desklib score. Measured (devtools/compare_detectors.py): AI samples 18/18 flagged, human abstracts from 2008 0/21, paragraphs our own humanizer had brought to 6%: 6/16 flagged (was 1/16). RADAR was tested and rejected (separation 0.72 on academic text). DETECTOR_MODEL3="" turns it off; DETECTOR_WEIGHT3 sets its share (0.5). Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>eebf465
9 days ago
-
Add meaning check, Groq model fallback and subject test run; close guard gaps
unknown +1: Claude Opus 5.5Found by running Smart Humanize on 16 subject samples (devtools/batch.py). - groqai: on a rate limit the request moves to the next model in GROQ_MODELS (limits are per model), then back when the limit has passed. - groqai.verify + humanizer: every rewritten sentence is read against the author's sentence; one that changed the meaning (moved p-value, reversed comparison, dropped method, "but" -> "so") gets the next version that passes, else the original (HUM_VERIFY, HUM_VERIFY_ALTS). - textparse.guard: integrity rules now apply to every version; damaged citations are repaired or the rewrite rejected; each citation exactly once; bracket spacing normalised before the checks. - basic_grammar no longer treats "(Mishra and Mishra, 2012)" as a doubled word. - Sentence splitting: initials and case names ("Justice K.S. Puttaswamy v. Union of India") no longer end a sentence; detector and humanizer share one rule. - pubrules: micro-prefixed and compound units (ug/m3, mmol/g, MP…689da70 -
Harden rewrite integrity, stop the final polish undoing humanization, add tests
unknown +1: Claude Opus 5.5Integrity (textparse, pubrules) - Guard rejects lost relation signs (p < 0.01 -> p 0.01), duplicated or invented citations and unbalanced brackets; closes sentences left open. - New rules: no added cause, no added "we"; mediate/moderate/correlate/ regress protected in any word form. - Fewer false rejections: "indicate" is a reporting verb, not a hedge; another form of an author's name is not invented. - Author's typography kept (no-break hyphens, curly quotes, minus sign); fragments keep their open ending and lower-case start. - strict_check() returns the rejection reason. Humanizer / Smart Humanize - Final polish repairs only sentences that fail the clarity checks; a clear humanized sentence is no longer reverted to the original, with a paragraph-level check on optional edits. - Rescue versions are written in four distinct manners and chosen next to their neighbours' current wording. - Word: bold/italic runs inside a rewritten sentence keep their format; no full stop …585f232
14 days ago
-
Upload detector.py
Verified signaturea4e3188
17 days ago
-
Upload 6 files
Verified signature12739bc
18 days ago
-
Upload 3 files
Verified signature1de19f4
20 days ago
-
Upload 4 files
Verified signatureef25ab8 -
Upload 6 files
Verified signaturebc8b4d1
21 days ago
-
Upload 4 files
Verified signature1f74b41 -
Upload 6 files
Verified signature5fa71d5 -
Upload 4 files
Verified signatureb4ef4c5
22 days ago
-
Upload 15 files
Verified signaturedd156e7 -
Upload 14 files
Verified signature41bc882 -
Upload 13 files
Verified signatureaa35e7b