TeenyTinyLlama-160m / README.md
nicholasKluge's picture
Update README.md
48c15a5
|
raw
history blame
8.71 kB
metadata
license: apache-2.0
datasets:
  - nicholasKluge/portuguese-corpus-v3
language:
  - pt
metrics:
  - perplexity
library_name: transformers
pipeline_tag: text-generation
tags:
  - text-generation-inference
widget:
  - text: Astronomia é uma ciência natural que estuda
    example_title: Exemplo
  - text: Em um achado chocante, o cientista descobriu um
    example_title: Exemplo
  - text: Python é uma linguagem de
    example_title: Exemplo
  - text: O Gato de Schrödinger é uma experiência mental
    example_title: Exemplo
inference:
  parameters:
    repetition_penalty: 1.5
    temperature: 0.3
    top_k: 30
    top_p: 0.3
    max_new_tokens: 200
co2_eq_emissions:
  emissions: 5.6
  source: CodeCarbon
  training_type: pre-training
  geographical_location: Germany
  hardware_used: NVIDIA A100-SXM4-40GB

Teeny-tiny-llama-162m (Portuguese)

A little llama wearing a mushroom hat and a monocle.

Teeny-tiny-llama-162m is a compact language model based on the Llama 2 architecture (Tiny-llama implementation). This model is designed to deliver efficient natural language processing capabilities (in Portuguese-BR) while being resource-conscious.

Teeny-tiny-llama has been trained by leveraging scaling laws to determine the optimal number of tokens per parameter while incorporating preference pre-training.

  • Compact Design: Teeny-tiny-llama is a downsized version of the Llama 2 architecture, making it suitable for applications with limited computational resources.

  • Optimized Scaling: The model has been pre-trained using scaling logs to identify the ideal token-to-parameter ratio.

  • Custom Portuguese Dataset: Teeny-tiny-llama has been trained on a custom Portuguese dataset. This dataset includes diverse linguistic contexts and preference pre-training, allowing the model to better cater to Portuguese language nuances and be better suited for fine-tuning tasks like instruction-tuning.

Details

  • Size: 162 million parameters
  • Dataset: Portuguese-Corpus-v3
  • Language: Portuguese
  • Number of steps: 457,969
  • Batch size: 4
  • Optimizer: torch.optim.AdamW (warmup_ratio = 0.01, learning_rate = 6e-4, epsilon = 1e-8)
  • GPU: 1 NVIDIA A100-SXM4-40GB
  • Training time: ~ 36 hours
  • Emissions: 5.6 KgCO2 (Germany)
  • Total Energy Consumption: 15.5 kWh

This repository has the source code used to train this model.

Training Set-up

Section Setting Value
Model args. vocab_size 32000
hidden_size 768
intermediate_size 3072
max_position_embeddings 2048
num_attention_heads 12
num_hidden_layers 12
num_key_value_heads 12
torch_dtype "float32"
Data args. dataset_name "nicholasKluge/portuguese-corpus-v3"
dataset_split "train"
train_num_samples 1831873
val_num_samples 18000
block_size 2048
Training args. evaluation_strategy "steps"
eval_steps 100000
per_device_train_batch_size 4
per_device_eval_batch_size 4
gradient_accumulation_steps 1
learning_rate 0.0006
adam_epsilon 0.00000001
weight_decay 0.01
lr_scheduler_type "cosine"
warmup_ratio 0.01
num_train_epochs 1
gradient_checkpointing false
seed 42
mixed_precision 'no'
checkpointing_steps 22000
tf32 true

Usage

# Use a pipeline as a high-level helper
from transformers import pipeline

pipe = pipeline("text-generation", model="nicholasKluge/Teeny-tiny-llama-162m")

# Load model directly
from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("nicholasKluge/Teeny-tiny-llama-162m")
model = AutoModelForCausalLM.from_pretrained("nicholasKluge/Teeny-tiny-llama-162m")

Limitations

🤥 Generative AI models, like LLMs used for text generation/conversation or GANs for image generation, can produce content that can be mistaken for truth but is, in fact, misleading or entirely false, given the model's tendency to output hallucinations. Such models can generate deceptive visuals, human-like textual content, music, or combined media that might seem genuine at first glance.

🤬 Machine learning systems can inherit social and historical stereotypes from the data used to train them. Given these biases, models can be prone to produce toxic content, that is, text, images, videos, or comments, that is harmful, offensive, or detrimental to individuals, groups, or communities. Also, models that automate decision-making can have biases against certain groups, affecting people based on sensitive attributes in an unjust manner.

Evaluations

Steps Evaluation Loss Perplexity Total Energy Consumption Emissions
100.000 3.19 24.52 3.75 kWh 1.28 CO2eq
200.000 3.02 20.58 7.51 kWh 2.56 CO2eq
300.000 2.83 16.98 11.25 kWh 3.84 CO2eq
400.000 2.79 16.41 14.52 kWh 5.11 CO2eq

Benchmarks

Models Average ARC Hellaswag MMLU TruthfulQA
Gpt2-portuguese-small 30.22 22.48 29.62 27.36 41.44

Cite as 🤗


@misc{nicholas22llama,
  doi = {10.5281/zenodo.6989727},
  url = {https://huggingface.co/nicholasKluge/Teeny-tiny-llama-162m},
  author = {Nicholas Kluge Corrêa},
  title = {Teeny-tiny-llama},
  year = {2023},
  publisher = {HuggingFace},
  journal = {HuggingFace repository},
}

License

The Teeny-tiny-llama-162m is licensed under the Apache License, Version 2.0. See the LICENSE file for more details.