Request access to mitra D3
Access is reviewed by hand and granted for research only.
This artifact belongs to a research system that rewrites LLM-written text in the style of named human writers, so that AI-text detectors may judge it human. Access is granted for research on AI-text detection, robustness of detectors, and style transfer. By requesting access you agree (1) not to present text produced with it as human-written work (academic, journalistic, commercial or any other setting), (2) not to impersonate the writers whose styles it imitates or any other person, (3) not to use it to deceive readers or detection systems outside a research evaluation, and (4) not to redistribute it.
Log in or Sign Up to review the conditions and access this model content.
mitra D3: own human-vs-LLM passage detector
Access and terms (gated, manual approval)
This is a public, gated research release, and access is approved by hand. It is not cleared for general redistribution:
- the training targets include copyrighted human writing (personal blogs, Hacker News comments, all pre-2022-11-30);
- it is part of a system that imitates named, real writers and whose outputs are built to be judged human-written by AI-text detectors.
Use it only for research on detection and style transfer. Never use it to pass AI text off as human work, to impersonate anyone, or to deceive. Do not redistribute the weights.
It is a component of the mitra rewriter, not a general AI-text detector. It was trained on human passages (blogs, Hacker News) and their LLM paraphrases.
DeBERTa-v3-small (revision a36c739020e01763fe789b4b85e2df55d6180012) fine-tuned on our own labels only:
- a human passage (1) vs a Muse (spark-1.3) paraphrase of that same passage (0);
- 9,846 train pairs, max-len 512, passages;
- dev pairwise accuracy 1.0 at step 400.
No Pangram verdict was used in training or selection.
In ampixa/mitra it picks, among a round's candidates that are close enough in
meaning to the source, the one it rates most human (train_hip rewrite --detector). It reads a document as greedy
windows of ≤ 512 tokens, scored as the token-weighted mean or the weakest window (--detector-agg min).
Known limits (S-F63):
- It is saturated on the RL models it was used to train. 92% of S37's candidates score above 6 log-odds.
- As a diagnostic on frozen S37 outputs, its AUC against Pangram's human verdict is 0.535 (mean of windows) and 0.567 (weakest window).
- A refreshed detector (D5, which adds our models' own outputs and two more LLM families) replaces it if D5 passes the project's gates.
Layout: manifest.json, complete.json, best/ (model + tokenizer + step.json).
Model tree for ampixa/mitra-detector-d3
Base model
microsoft/deberta-v3-small