Table 2. Dataset sizes.
The human text datasets used in our experiments were taken from the samples of human text published with the respective language models and resized to match the size of the machine text datasets.
| Model | Dataset full name | Short | Full | Filtered | ||||
|---|---|---|---|---|---|---|---|---|
| Name | Train | Valid | Test | Train | Valid | Test | ||
| Machine datasets | ||||||||
| GPT2 | Small-117M | s | 250,000 | 5,000 | 5,000 | 185,622 | 3,732 | 3,722 |
| GPT2 | xl-1542M | xl | 250,000 | 5,000 | 5,000 | 193,052 | 3,868 | 3,851 |
| GPT2 | Small-117M-k40 | s-k | 250,000 | 5,000 | 5,000 | 201,236 | 4,062 | 4,082 |
| GPT2 | xl-1542M-k40 | xl-k | 250,000 | 5,000 | 5,000 | 214,202 | 4,312 | 4,243 |
| GPT3 | 175B | GPT3 | 1,604 | 201 | 201 | 886 | 122 | 101 |
| Grover | Grover-Mega | Grover | 8,000 | 1,000 | 1,000 | 7,740 | 964 | 961 |
| Human datasets | ||||||||
| GPT2 | Webtext | 250,000 | 5,000 | 5,000 | 190,503 | 3,813 | 3,834 | |
| GPT3 | GPT3-webtext | 1,604 | 201 | 201 | 1,235 | 160 | 155 | |
| Grover | realNews | 8,000 | 1,000 | 1,000 | 7,725 | 972 | 976 | |