lm-gpt2-timemachine

This model is a causal language model trained from scratch on the novel
The Time Machine by H. G. Wells.

The objective of this project is educational: to understand the full pipeline of training a GPT-style language model from raw text, including preprocessing, tokenization, training, and text generation.

---

Model description

• Architecture: GPT-2–style causal language model
• Training type: From scratch (no pretrained weights)
• Language: English
• Tokenizer: Byte-level BPE
• Context length: 128 tokens
• Task: Causal language modeling (next-token prediction)

The model learns to predict the next token given a sequence of previous tokens, and can be used to generate text in the style of The Time Machine.

---

Intended uses & limitations

Intended uses

• Educational purposes
• Learning how causal language models work
• Small-scale text generation experiments
• Understanding Hugging Face training workflows

Limitations

• Trained on a very small corpus (single novel)
• Not suitable for general-purpose text generation
• May produce repetitive or incoherent text
• Not optimized for factual correctness

---

Training and evaluation data

• Dataset: The Time Machine by H. G. Wells
• Source: Public domain text
• Preprocessing:
  • Text normalization
  • Tokenization with a byte-level tokenizer
  • Concatenation of text and splitting into fixed-length blocks (128 tokens)

The same dataset was split into training, validation, and test subsets.

---

Training procedure

• Objective: Causal Language Modeling
• Loss function: Cross-entropy loss
• Optimizer: AdamW
• Batching: Fixed-length token blocks
• Evaluation: Perplexity on validation data

The model was trained using the Hugging Face Trainer API.

---

Example generation

The Time Traveller (for so it will be convenient to speak of him) was a curious man,
of no less intellectual character than...


### Training hyperparameters

The following hyperparameters were used during training:
- learning_rate: 5e-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: 3
- mixed_precision_training: Native AMP

### Training results

- 'eval_loss': 2.851572036743164
- 'eval_perplexity': 17.314979553222656

### Framework versions

- Transformers 4.57.6
- Pytorch 2.6.0+cu126
- Datasets 3.6.0
- Tokenizers 0.22.1