---
license: mit
language:
- en
pipeline_tag: text-classification
base_model: Oxidane/tmr-ai-text-detector
base_model_relation: finetune
tags:
- onnx
- roberta
- ai-text-detection
- onnxruntime
- wasm
- browser
---

# Draftmark RoBERTa FT INT8

Draftmark is a free, on-device writing-pattern detector for English prose. This is the exact weight-only INT8 ONNX export used by the live browser release. Its model and tokenizer assets total **131.1 MB**. [Live Draftmark](https://draftmark-virid.vercel.app/) · [Release manifest](release-manifest.json)

## Inference and intended use

The ONNX graph at `onnx/model.int8.onnx` accepts `input_ids` and `attention_mask` as rank-two INT64 tensors. Its two logits are ordered `human`, then `ai`. Apply a softmax per window. The shipped worker scores sentence-aligned windows of up to 512 tokens including special tokens and averages the AI-class scores weighted by content-token count. Oversized sentences are split across windows. The shipped flag threshold is **0.999146566314596** and the limited-signal cutoff is **0.25**.

The evaluated app accepts English prose of 150–2,000 words and up to 20,000 characters. Scores can support manual review. Authorship decisions require independent evidence. Results depend on the generator, rewriting method, domain, and writing population.

## Fresh holdout results

| Measure | Previous Ettin | Draftmark FT INT8 |
| --- | ---: | ---: |
| Unedited AI recall | 215/300 | 277/300 |
| Decoder rewrite recall | 18/150 | 89/150 |
| Sequence-to-sequence rewrite recall | 132/150 | 110/150 |
| Mean paraphrase recall | 50.0% | 66.3% |
| General-human false flags | 10/300 | 1/300 |
| Learner-essay false flags | 80/220 | 0/220 |

The paired 95% confidence interval for the mean paraphrase gain was **+10.7 to +22.0 percentage points**. Sequence-to-sequence recall declined. The learner false-flag count applies to the 220 ASAP 2.0 essays evaluated. The results are a comparison against Draftmark’s previous detector on these samples. [Aggregate benchmark](finetune-v3.json)

## Training and attribution

Draftmark fine-tuned the last four encoder layers and the classification head of Oxidane’s MIT-licensed TMR RoBERTa at revision `0ceddea903015ef99cbaa040a4d8a216aed9c683`. Training used source-grouped DetectRL-X pairs, original AI and human prose from DetectionAI, and learner essays from the ASAP 2.0 corpus and the Shijaku–Canhasi TOEFL collection. The report records the source pins, split sizes, and overlap screening. The ASAP 2.0 corpus is CC BY 4.0 and is credited to Scott Crossley and contributors, Vanderbilt University, and The Learning Agency Lab. [TMR](https://huggingface.co/Oxidane/tmr-ai-text-detector) · [DetectRL-X](https://aclanthology.org/2026.acl-long.1773/) · [DetectionAI](https://github.com/brianjabarian/DetectionAI) · [ASAP 2.0](https://github.com/scrosseye/ASAP_2.0) · [Shijaku–Canhasi corpus](https://github.com/rexshijaku/chatgpt-generated-text-detection-corpus)

The [GPU handoff](handoff.md) documents how to reconstruct a Transformers checkpoint from the pinned base and the [replacement tensors on the training branch](https://huggingface.co/saksornr/draftmark-roberta-int8/blob/training/training/trainable.safetensors). The main branch is the browser distribution.