cglez/gpt2-ag_news

Model Card: GPT-2-AG-News

An in-domain GPT-2, pre-trained from scratch on the AG-News dataset texts.

Model Details

Description

This model is based on the GPT-2 architecture and was pre-trained from scratch (in-domain) using the text in AG-News dataset, excluding its test split.

• Developed by: Cesar Gonzalez-Gutierrez
• Funded by: ERC
• Architecture: GPT-2
• Language: English
• License: MIT
• Base model: GPT-2

Checkpoints

Intermediate checkpoints from the pre-training process are available and can be accessed using specific tags, which correspond to training epochs and steps:

Epoch	Step	Tags
1	1125	epoch-1	step-1125
5	5625	epoch-5	step-5625
10	11250	epoch-10	step-11250
20	22500	epoch-20	step-22500
30	33750	epoch-30	step-33750
40	45000	epoch-40	step-45000
50	56250	epoch-50	step-56250
60	67500	epoch-60	step-67500
70	78750	epoch-70	step-78750
80	90000	epoch-80	step-90000
90	101250	epoch-90	step-101250
100	112500	epoch-100	step-112500

To load a model from a specific intermediate checkpoint, use the revision parameter with the corresponding tag:

from transformers import AutoModelForCausalLM

model = AutoModelForMaskedLM.from_pretrained("<model-name>", revision="<checkpoint-tag>")

Sources

• Paper: [Information pending]

Training Details

For more details on the training procedure, please refer to the base model's documentation: Training procedure.

Training Data

All texts from AG-News dataset, excluding the test partition.

Training Hyperparameters

• Precision: fp16
• Batch size: 8
• Gradient accumulation steps: 12

Uses

For typical use cases and limitations, please refer to the base model's guidance: Inteded uses & limitations.

Bias, Risks, and Limitations

This model inherits potential risks and limitations from the base model. Refer to: Limitations and bias.

Environmental Impact

• Hardware Type: NVIDIA A100 PCIE 40GB
• Hours used: 15
• Cluster Provider: Artemisa
• Compute Region: EU
• Carbon Emitted: 1.62 kg CO2 eq.

Citation

BibTeX:

[More Information Needed]