Moodlerz/deberta-v3-detector-eli5

What is this?

This model was fine-tuned as part of a research project comparing transformer-based AI-text detectors across two benchmark datasets: HC3 and ELI5.

The task is binary classification:

• Label 0 → Human-written text
• Label 1 → LLM-generated text

Model details

DeBERTa-v3-base (microsoft/deberta-v3-base) fine-tuned on ELI5 for AI-text detection. Binary classifier: human (0) vs LLM-generated (1). Trained in full fp32 (bf16/fp16 disabled due to disentangled attention gradient issues). No intermediate checkpointing — final in-memory weights used directly.

Training setup

Setting	Value
Epochs	1
Batch size (train)	16
Learning rate	2e-5
Warmup steps	500
Weight decay	0.01
Dropout	0.2
Max seq length	512
Validation split	10%
Best model metric	ROC-AUC

Datasets

• HC3 — Human ChatGPT Comparison Corpus
• ELI5 — Explain Like I'm 5 (Reddit QA dataset)
  Cross-dataset evaluation (e.g. trained on HC3, tested on ELI5) was used to measure generalisability of each detector.

How to load

from transformers import AutoTokenizer, AutoModelForSequenceClassification
import torch

model = AutoModelForSequenceClassification.from_pretrained("Moodlerz/deberta-v3-detector-eli5")
tokenizer = AutoTokenizer.from_pretrained("Moodlerz/deberta-v3-detector-eli5")

text = "Your input text here"
inputs = tokenizer(text, return_tensors="pt", truncation=True, max_length=512)
with torch.no_grad():
    logits = model(**inputs).logits
prob_llm = torch.softmax(logits, dim=-1)[0][1].item()
print(f"P(LLM-generated): {prob_llm:.4f}")

Notes

• Local training dir: ./models/DeBERTa_eli5
• All models in this series are private repos under Moodlerz.
• Part of a larger study — do not use for production content moderation without further evaluation.