Greetings & Wishes Generator - educational group project as part of the course on neural networks (NLP) at ITMO. We have fine tuned the rugpt3small_based_on_gpt2 model from sber-ai on the data that we collected from the site.
The result we got:
***** train metrics *****
epoch = 3.0
total_flos = 1312249GF
train_loss = 1.8419
train_runtime = 0:25:57.44
train_samples = 3595
train_samples_per_second = 6.925
train_steps_per_second = 6.925
***** eval metrics *****
epoch = 3.0
eval_accuracy = 0.6166
eval_loss = 1.971
eval_runtime = 0:00:23.57
eval_samples = 913
eval_samples_per_second = 38.734
eval_steps_per_second = 38.734
perplexity = 7.1779
Mean Cosine Similarity: 0.6579082608222961
.
├───data
│ ├───raw
│ │ └───greetings.csv # csv file with all greetings & wishes
│ ├───processed
│ └───├───train.txt # dataset for training
│ └───valid.txt # dataset for validation
│
├───src
│ └───parse.py # scrapes greetings & wishes from the website
├───.gitignore
├───Greetings_and_Wishes_Generator.ipynb # trains the model on the training dataset,
│ contains class with the tuned model,
│ example how the inference works
└───README.md
- download finetuning model;
- copy Greetings_and_Wishes_Generator.ipynb to your Google Drive and run the last partition with your data.