File size: 2,832 Bytes
79ed695
 
2c1556a
 
79ed695
 
 
 
 
 
 
e7bc810
79ed695
 
 
 
 
e7bc810
79ed695
 
 
 
 
 
 
e7bc810
79ed695
e7bc810
79ed695
 
 
 
 
 
 
bbbcf03
 
 
79ed695
 
 
bbbcf03
 
 
 
 
 
79ed695
 
 
bbbcf03
79ed695
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
---
language:
- en
- fr
license: mit
tags:
- lm-detection
datasets:
- hc3_multi_custom_ms_hg
metrics:
- f1
base_model: xlm-roberta-base
model-index:
- name: xlmr-chatgptdetect-noisy
  results:
  - task:
      type: text-classification
      name: Text Classification
    dataset:
      name: HC3 FULL_MULTI_1.0_0.5_0.5
      type: glue
      config: full_multi_1.0_0.5_0.5
      split: vsl
      args: full_multi_1.0_0.5_0.5
    metrics:
    - type: f1
      value: 0.963274059512108
      name: F1
---

# xlmr-chatgptdetect-noisy

Multilingual ChatGPT detection model from [Towards a Robust Detection of Language Model-Generated Text: Is ChatGPT that easy to detect?](TODO:)

This model is a fine-tuned version of [xlm-roberta-base](https://huggingface.co/xlm-roberta-base) on the HC3 FULL_MULTI_1.0_0.5_0.5 dataset with noise added.
It achieves the following results on the:

Evaluation set:
- Loss: 0.1573
- F1: 0.9633

Test Set:
- F1: 0.97

Adversarial:
- F1: 0.45

## Model description

This a model trained to detect text created by ChatGPT in French.
The training data is the combination of the `hc3_fr_full` and `hc3_en_full` subsets of [almanach/hc3_multi](https://huggingface.co/datasets/almanach/hc3_french_ood), but with added misspelling and homoglyph attacks.

## Intended uses & limitations

This model is for research purposes only.
It is not intended to be used in production as we said in our paper:

**We would like to emphasize that our study does not claim to have produced an universally accurate detector. Our strong results are based on in-domain testing and, unsurprisingly, do not generalize in out-of-domain scenarios. This is even more so when used on text specifically designed to fool language model detectors and on text intentionally stylistically similar to ChatGPT-generated text, especially instructional text.**

## Training procedure

### Training hyperparameters

The following hyperparameters were used during training:
- learning_rate: 2e-05
- train_batch_size: 8
- eval_batch_size: 8
- seed: 1
- gradient_accumulation_steps: 4
- total_train_batch_size: 32
- optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
- lr_scheduler_type: linear
- lr_scheduler_warmup_ratio: 0.1
- num_epochs: 5.0
- mixed_precision_training: Native AMP

### Training results

| Training Loss | Epoch | Step  | Validation Loss | F1     |
|:-------------:|:-----:|:-----:|:---------------:|:------:|
| 0.0317        | 1.0   | 8538  | 0.1732          | 0.9492 |
| 0.008         | 2.0   | 17076 | 0.3541          | 0.9270 |
| 0.0085        | 3.0   | 25614 | 0.1161          | 0.9726 |
| 0.0015        | 4.0   | 34152 | 0.2557          | 0.9516 |
| 0.0           | 5.0   | 42690 | 0.2286          | 0.9650 |


### Framework versions

- Transformers 4.26.1
- Pytorch 1.11.0+cu115
- Datasets 2.8.0
- Tokenizers 0.13.2