gpt2-wikitext2

This model is a fine-tuned version of gpt2 on an wikitext2 dataset. It achieves the following results on the evaluation set:

• Loss: 6.3377

Model description

More information needed

Intended uses & limitations

More information needed

Training and evaluation data

More information needed

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 2e-05
• train_batch_size: 8
• eval_batch_size: 8
• seed: 42
• optimizer: Use OptimizerNames.ADAMW_TORCH with betas=(0.9,0.999) and epsilon=1e-08 and optimizer_args=No additional optimizer arguments
• lr_scheduler_type: linear
• num_epochs: 2.0

Training results

Training Loss	Epoch	Step	Validation Loss
6.3991	1.0	2249	6.5169
6.2508	2.0	4498	6.3377

Framework versions

• Transformers 4.52.4
• Pytorch 2.6.0+cu124
• Datasets 3.6.0
• Tokenizers 0.21.2

Language Model Training Notebook

• Causal Language Modeling (CLM): Training a model to predict the next token in a sequence. (Current model)
• Masked Language Modeling (MLM): Training a model to predict masked tokens in a sequence.

Dataset

I used the Wikitext 2 dataset as an example in this notebook, but you can easily adapt it to use other datasets from the Hugging Face Hub or your own local data.

Pre-training

I fine-tuned a GPT-2 model using the steps outlined in this notebook.

• Model Architecture: I used the gpt2 model checkpoint.
• Dataset: I used the wikitext-2-raw-v1 dataset.
• Training Duration: I trained for 2 epochs.
• Key Configurations:
  • Learning rate: 2e-5
  • Weight decay: 0.01
  • Batch size: Determined by per_device_train_batch_size in TrainingArguments (default is 8). Gradient accumulation steps are not set (default is 1).
  • Optimizer: AdamW (default in Trainer)
  • Scheduler: Linear with warmup (default in Trainer)
  • Checkpointing: Saved checkpoint every epoch, keeping the last 2.

GPT-2 Architecture

The GPT-2 model is based on a decoder-only transformer architecture. Key architectural details:

• Type: Decoder-only Transformer
• Layers: 12 transformer blocks (GPT-2 base)
• Hidden Size: 768
• Attention Heads: 12
• Vocabulary Size: 50,257
• Max Sequence Length: 1024 tokens
• Positional Embeddings: Learned
• Activation Function: GELU
• Layer Normalization: Applied before attention and feed-forward layers (pre-LN)
• Causal Self-Attention: Each token attends only to previous tokens