igorktech/CharPicoSatirik-sm

TinySatirik-sm

This model is a pre-trained version of really tiny LLama2 model on an anekdots dataset.

Inspired by TinyStories.

It achieves the following results on the evaluation set:

• Loss: 1.2643

Tokenizer

To utilize the model, install the special tokenizer:

pip install git+https://github.com/Koziev/character-tokenizer

In addition to recognizing Cyrillic characters and punctuation, this tokenizer is aware of special tokens such as <s>, </s>, <pad>, and <unk>.

As this is a non-standard tokenizer for transformers, load it not via transformers.AutoTokenizer.from_pretrained, but somewhat like this:

import charactertokenizer

...
tokenizer = charactertokenizer.CharacterTokenizer.from_pretrained('igorktech/CharPicoSatirik-sm')

To observe tokenization, use this code snippet:

prompt = '<s>Hello World\n'
encoded_prompt = tokenizer.encode(prompt, return_tensors='pt')
print('Tokenized prompt:', ' | '.join(tokenizer.decode([t]) for t in encoded_prompt[0]))

You will see a list of tokens separated by the | symbol:

Tokenized prompt: <s> | H | e | l | l | o |   | W | o | r | l | d | 

Tokenizer created by Koziev.

Model description

Llama2 architecture based.

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: 0.0005
• train_batch_size: 32
• eval_batch_size: 2
• seed: 42
• gradient_accumulation_steps: 4
• total_train_batch_size: 128
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: cosine
• lr_scheduler_warmup_steps: 250
• num_epochs: 5
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss
1.3401	1.81	2000	1.3465
1.2323	3.62	4000	1.2643

Framework versions

• Transformers 4.36.0.dev0
• Pytorch 2.1.0+cu121
• Datasets 2.16.1
• Tokenizers 0.15.0