glyph-v1.1 / README.md
ogmatrixllm's picture
Update README.md
105c9ff verified
---
license: mit
language:
- en
tags:
- text-classification
- ai-text-detection
- deberta-v3
- binary-classification
- nlp
datasets:
- liamdugan/raid
- artem9k/ai-text-detection-pile
- gsingh1-py/train
- cc_news
- blog_authorship_corpus
- webis/tldr-17
- ChristophSchuhmann/essays-with-instructions
- HuggingFaceH4/stack-exchange-preferences
- pile-of-law/pile-of-law
metrics:
- accuracy
- f1
- precision
- recall
- roc_auc
pipeline_tag: text-classification
model-index:
- name: GLYPH
results:
- task:
type: text-classification
name: AI-Generated Text Detection
metrics:
- name: Accuracy
type: accuracy
value: 0.9885
- name: F1
type: f1
value: 0.9901
- name: Precision
type: precision
value: 0.9851
- name: Recall
type: recall
value: 0.9952
- name: ROC-AUC
type: roc_auc
value: 0.9990
- name: MCC
type: mcc
value: 0.9765
---
# GLYPH — High-Accuracy AI Text Detector
GLYPH is a binary text classifier built on [DeBERTa-v3-base](https://huggingface.co/microsoft/deberta-v3-base) that distinguishes human-written text from AI-generated text. It achieves **98.85% accuracy**, **0.999 ROC-AUC**, and **0.990 F1** on a held-out test set spanning 10 human writing domains and 14 AI model families — from GPT-2 (1.5B) through GPT-4 (~1T).
The model was trained on ~50K texts covering academic papers, news articles, blog posts, Reddit discussions, legal filings, Wikipedia, student essays, and technical Q&A on the human side, and outputs from 24 distinct AI model configurations across 10 model families on the AI side. It produces well-separated, high-confidence predictions (mean confidence 0.976) and remains accurate even at the strictest decision thresholds.
## Key Results
| Metric | Value |
|---|---|
| **Accuracy** | 98.85% |
| **F1 Score** | 0.9901 |
| **Precision** | 98.51% |
| **Recall** | 99.52% |
| **ROC-AUC** | 0.9990 |
| **Average Precision** | 0.9993 |
| **MCC** | 0.9765 |
| **Human Accuracy** | 97.94% |
| **AI Accuracy** | 99.52% |
| **Mean Confidence** | 0.976 |
| **F1 @ 0.95 threshold** | 0.987 |
All metrics evaluated on a held-out test set of 5,050 texts (2,136 human / 2,914 AI) with no overlap in source texts, split hashes, or temporal leakage with the training set.
## Per-Source Performance
### Human Text Sources
| Source | Domain | n | Accuracy | Confidence |
|---|---|---|---|---|
| PubMed Abstracts | Biomedical research | 300 | **100.0%** | 0.988 |
| Blog / Opinion | Personal blogs | 200 | **100.0%** | 0.987 |
| Reddit Writing | Informal / social | 300 | **100.0%** | 0.985 |
| Wikipedia | Encyclopedic | 500 | **99.8%** | 0.987 |
| CC-News | Journalism | 392 | **99.5%** | 0.981 |
| arXiv Abstracts | Academic / scientific | 444 | **90.8%** | 0.948 |
arXiv abstracts are the hardest category — highly formulaic academic prose with structural similarity to AI output. Even so, detection accuracy is 90.8% with 94.8% mean confidence, and the remaining errors are concentrated in a small subset of unusually short or template-heavy abstracts.
### AI Model Families
| Model | Family | Params | n | Accuracy | F1 |
|---|---|---|---|---|---|
| GPT-3.5-Turbo | OpenAI | 175B | 223 | **100.0%** | 1.000 |
| GPT-4 | OpenAI | ~1T | 215 | **100.0%** | 1.000 |
| Llama-2-70B-Chat | Meta | 70B | 191 | **100.0%** | 1.000 |
| MPT-30B | MosaicML | 30B | 211 | **100.0%** | 1.000 |
| MPT-30B-Chat | MosaicML | 30B | 191 | **100.0%** | 1.000 |
| Mistral-7B-Instruct-v0.1 | Mistral AI | 7B | 194 | **100.0%** | 1.000 |
| Mistral-7B-v0.1 | Mistral AI | 7B | 203 | **100.0%** | 1.000 |
| Llama-3.1-8B-Instruct | Meta | 8B | 238 | **99.6%** | 0.998 |
| Phi-3.5-Mini-Instruct | Microsoft | 3.8B | 238 | **99.6%** | 0.998 |
| Command-Chat | Cohere | 52B | 198 | **99.5%** | 0.997 |
| Text-Davinci-002 | OpenAI | 175B | 176 | **99.4%** | 0.997 |
| Llama-3.2-3B-Instruct | Meta | 3B | 238 | **99.2%** | 0.996 |
| GPT-2-XL | OpenAI | 1.5B | 198 | **98.5%** | 0.992 |
| Cohere Command | Cohere | 52B | 200 | **97.5%** | 0.987 |
Detection is robust across four generations of language models (GPT-2 through GPT-4), three access paradigms (open-weight, API-only, and proprietary), and parameter counts spanning three orders of magnitude (1.5B to ~1T).
### Performance by Text Length
| Length Bucket | n | Accuracy | F1 |
|---|---|---|---|
| Very Long (>2000 words) | 103 | **100.0%** | 1.000 |
| Long (500–2000 words) | 862 | **99.9%** | 0.999 |
| Short (50–150 words) | 1,976 | **98.5%** | 0.989 |
| Medium (150–500 words) | 1,634 | **98.8%** | 0.989 |
| Very Short (<50 words) | 475 | **98.1%** | 0.899 |
Performance degrades gracefully with shorter inputs. Even on texts under 50 words — where the model has minimal signal — accuracy remains above 98%.
### Threshold Sensitivity
The model produces well-calibrated, high-confidence outputs. Performance holds across aggressive decision thresholds:
| P(AI) Threshold | F1 | Precision |
|---|---|---|
| 0.50 (default) | 0.990 | 0.985 |
| 0.60 | 0.991 | 0.987 |
| 0.70 | 0.992 | 0.990 |
| 0.80 | 0.992 | 0.992 |
| 0.90 | 0.991 | 0.993 |
| 0.95 | 0.987 | 0.996 |
At a 0.95 threshold, precision reaches 99.6% with only a 0.3% drop in F1 — suitable for high-stakes applications where false accusations of AI usage carry serious consequences.
## Architecture
| Component | Details |
|---|---|
| Base model | `microsoft/deberta-v3-base` (184M parameters) |
| Architecture | DeBERTa-v3 with disentangled attention and enhanced mask decoder |
| Task head | Linear classifier (768 → 2) with 0.15 dropout |
| Tokenizer | SentencePiece (slow tokenizer, `use_fast=False`) |
| Max sequence length | 512 tokens |
| Output | `[P(human), P(AI)]` softmax probabilities |
DeBERTa-v3 was chosen over RoBERTa and BERT alternatives due to its disentangled attention mechanism, which separately encodes content and position. This is particularly relevant for AI text detection: language models have characteristic positional dependencies in how they distribute tokens across a sequence, and disentangled attention gives the classifier direct access to these patterns.
## Training
### Configuration
| Parameter | Value |
|---|---|
| Trainable parameters | 184,423,682 (100% — all layers unfrozen) |
| Optimizer | AdamW (weight decay 0.01) |
| Learning rate | 2e-5 (cosine schedule) |
| Warmup | 10% of total steps |
| Effective batch size | 64 (16 × 4 gradient accumulation) |
| Precision | bf16 mixed precision |
| Gradient checkpointing | Enabled (non-reentrant) |
| Label smoothing | 0.05 |
| Class weights | human=1.182, ai=0.867 |
| Epochs | 8 (early-stopped at 3.17) |
| Best checkpoint | Epoch 1.19 (by validation F1) |
| Training time | ~49 minutes on RTX 4070 Ti 12GB |
| Final train loss | 0.186 |
| Final eval loss | 0.150 |
### Why Fully Unfrozen?
Initial experiments with 4 frozen encoder layers (standard practice from PAN-CLEF 2025 literature) yielded only 80% accuracy with severe human-side bias — the model classified 44% of human texts as AI. Freezing 4 of 12 layers in DeBERTa-base locks 33% of the network, far more aggressive than the 21% reported for DeBERTa-large. Unfreezing all layers with cosine LR decay and 10% warmup resolved the bias entirely, lifting human accuracy from 55.6% to 97.9% without sacrificing AI detection (97.4% → 99.5%).
### Dataset Composition
**Total: 50,458 texts** (40,364 train / 5,044 validation / 5,050 test)
Stratified by source with hash-based deduplication to prevent data leakage.
#### Human Sources (10 domains, ~29K target)
| Domain | Source | Target Count | Text Type |
|---|---|---|---|
| Academic (STEM) | arXiv API | 5,000 | Abstracts across 8 categories (cs.CL, cs.AI, cs.LG, physics, math, q-bio, econ, stat) |
| Academic (Medical) | PubMed API | 3,000 | Biomedical research abstracts |
| Encyclopedic | Wikipedia API | 5,000 | Article sections across 10 topic categories |
| Journalism | CC-News (HuggingFace) | 4,000 | News articles |
| Literary / Creative | Project Gutenberg | 2,000 | Public domain book excerpts |
| Informal / Social | Reddit (webis/tldr-17) | 3,000 | Writing-focused subreddit posts |
| Student / Educational | PERSUADE corpus | 2,000 | Student essays |
| Technical / Q&A | StackExchange | 2,000 | Technical answers |
| Blog / Opinion | Blog Authorship Corpus | 2,000 | Personal blog posts |
| Legal / Formal | Pile of Law | 1,000 | Legal opinions and case summaries |
#### AI Sources (24 model configurations across 10 families)
**Locally generated via LM Studio (8 models, Q4_K_M quantization):**
| Model | Family | Parameters |
|---|---|---|
| Llama-3.1-8B-Instruct | Meta Llama | 8B |
| Llama-3.2-3B-Instruct | Meta Llama | 3B |
| Mistral-7B-Instruct-v0.3 | Mistral AI | 7B |
| Qwen2.5-7B-Instruct | Alibaba Qwen | 7B |
| Qwen2.5-14B-Instruct | Alibaba Qwen | 14B |
| Gemma-2-9B-Instruct | Google | 9B |
| Phi-3.5-Mini-Instruct | Microsoft | 3.8B |
| DeepSeek-V2-Lite-Chat | DeepSeek | 16B (MoE) |
Local generation used 4 temperature/sampling configurations (default, creative, precise, varied) across 6 prompt strategies (direct, continue, rewrite, expand, style_mimic, question_answer) with a system prompt enforcing natural human-like output — no markdown, no meta-commentary, no self-referential AI language.
**HuggingFace datasets (16 additional model families):**
| Dataset | Models Added | Reference |
|---|---|---|
| RAID (ACL 2024) | ChatGPT-3.5, GPT-4, GPT-3-Davinci, Cohere Command, Llama-2-70B-Chat, Mistral-7B-v0.1, Mixtral-8x7B, MPT-30B, GPT-2-XL | [liamdugan/raid](https://huggingface.co/datasets/liamdugan/raid) |
| AI Text Detection Pile | GPT-2/3/J/ChatGPT (mixed) | [artem9k/ai-text-detection-pile](https://huggingface.co/datasets/artem9k/ai-text-detection-pile) |
| NYT Multi-Model | GPT-4o, Yi-Large, Qwen-2-72B, Llama-3-8B, Gemma-2-9B, Mistral-7B | [gsingh1-py/train](https://huggingface.co/datasets/gsingh1-py/train) |
This combination ensures coverage of proprietary API models (GPT-3.5, GPT-4, GPT-4o, Cohere), large open models exceeding consumer GPU VRAM (Llama-2-70B, Qwen-2-72B, Mixtral-8x7B, Yi-Large), older architectures (GPT-2, GPT-3, GPT-J), and mixture-of-experts models (Mixtral, DeepSeek-V2-Lite). RAID data was filtered to non-adversarial generations only (`attack=="none"`) for training data quality.
## Usage
### With Transformers
```python
from transformers import AutoTokenizer, AutoModelForSequenceClassification
import torch
model_name = "ogmatrixllm/glyph" # Replace with your repo path
tokenizer = AutoTokenizer.from_pretrained(model_name, use_fast=False)
model = AutoModelForSequenceClassification.from_pretrained(model_name)
model.eval()
text = "Your text to classify here..."
inputs = tokenizer(text, return_tensors="pt", truncation=True, max_length=512)
with torch.no_grad():
logits = model(**inputs).logits
probs = torch.softmax(logits, dim=-1)
p_human, p_ai = probs[0].tolist()
label = "AI-generated" if p_ai > 0.5 else "Human-written"
confidence = max(p_human, p_ai)
print(f"{label} (confidence: {confidence:.1%})")
```
### With Pipeline
```python
from transformers import pipeline
detector = pipeline(
"text-classification",
model="ogmatrixai/glyph", # Replace with your repo path
tokenizer=AutoTokenizer.from_pretrained("ogmatrixai/glyph", use_fast=False),
)
result = detector("Your text here...")
print(result)
# [{'label': 'LABEL_1', 'score': 0.98}] # LABEL_0 = human, LABEL_1 = AI
```
### Important Notes
- **Tokenizer**: Always use `use_fast=False`. The fast tokenizer for DeBERTa-v3 has a confirmed regression in `transformers>=4.47` ([#42583](https://github.com/huggingface/transformers/issues/42583)) that crashes on load.
- **Max length**: The model was trained with `max_length=512`. Longer texts should be truncated or chunked with predictions aggregated.
- **Labels**: `LABEL_0` = human, `LABEL_1` = AI-generated.
## Limitations and Ethical Considerations
### Known Limitations
1. **English only.** GLYPH was trained exclusively on English text. Performance on other languages is untested and likely degraded.
2. **Training distribution.** The model has seen outputs from 24 specific AI model configurations. Novel architectures, heavily fine-tuned models, or future model families may evade detection. AI text detection is fundamentally adversarial — no static detector provides permanent robustness.
3. **arXiv abstracts remain the hardest domain** at 90.8% accuracy. Highly formulaic academic writing with rigid structural conventions shares surface features with AI-generated text. Users in academic integrity contexts should treat borderline predictions on scientific abstracts with appropriate caution.
4. **Short texts (<50 words)** have reduced F1 (0.899) despite high accuracy (98.1%). With minimal token-level signal, the model occasionally produces confident but incorrect predictions. For short-form content, consider requiring higher confidence thresholds.
5. **Adversarial attacks.** The training data includes only non-adversarial AI outputs. Paraphrasing attacks, homoglyph substitution, targeted prompt engineering, and watermark-removal techniques were not included. Dedicated adversarial robustness (e.g., RAID adversarial subsets) is a planned enhancement.
6. **Mixed authorship.** GLYPH classifies at the document level. It does not detect partial AI usage (e.g., AI-written paragraphs embedded in a human-written essay). Sentence-level or span-level detection requires a different approach.
7. **512-token window.** Texts are truncated at 512 tokens. For long documents, this means classification is based on the opening ~350–400 words only. Sliding-window aggregation is recommended for long-form content.
### Ethical Considerations
AI text detection carries real consequences — academic penalties, professional reputation damage, content moderation decisions. False positives (human text classified as AI) are particularly harmful. While GLYPH's false positive rate is low (2.06% on the test set, 44 out of 2,136 human texts), no detector achieves zero false positives.
**Recommendations for responsible deployment:**
- Never use GLYPH as the sole basis for punitive action. Use it as one signal among many (metadata, behavioral patterns, stylometric analysis).
- Apply a high confidence threshold (≥0.95) for consequential decisions. At this threshold, precision reaches 99.6%.
- Provide users with the confidence score, not just a binary label. A text scored at P(AI)=0.52 is fundamentally different from one scored at P(AI)=0.99.
- Maintain an appeals process. Statistical classifiers will always produce errors.
- Acknowledge the base rate problem. In populations where AI usage is rare, even a 2% FPR produces many false accusations relative to true detections.
## Training Infrastructure
| Component | Specification |
|---|---|
| GPU | NVIDIA GeForce RTX 4070 Ti (12GB VRAM) |
| CPU | Intel Core i7-14700K (20 cores) |
| RAM | 48GB DDR5 |
| Framework | PyTorch 2.6+ / HuggingFace Transformers |
| Precision | bf16 mixed precision |
| Total training time | 49 minutes |
| Experiment tracking | Weights & Biases |
## Citation
```bibtex
@misc{glyph2026,
title={GLYPH: High-Accuracy AI Text Detection with DeBERTa-v3},
author={OGMatrix},
year={2026},
url={https://huggingface.co/ogmatrixllm/glyph}
}
```
## Acknowledgments
Training data incorporates the [RAID benchmark](https://huggingface.co/datasets/liamdugan/raid) (Dugan et al., ACL 2024), the [AI Text Detection Pile](https://huggingface.co/datasets/artem9k/ai-text-detection-pile), and the [NYT Multi-Model dataset](https://huggingface.co/datasets/gsingh1-py/train). Human text sources include arXiv, PubMed, Wikipedia, CC-News, Project Gutenberg, Reddit, StackExchange, Blog Authorship Corpus, PERSUADE, and Pile of Law. The base model is [DeBERTa-v3-base](https://huggingface.co/microsoft/deberta-v3-base) by Microsoft Research.