pinaltech/echo-replies
A model for quick reply suggestions
1 min read
The model is designed to automatically select short reply options for incoming messages. It analyzes the message text, determines its language and context, and then selects the most appropriate phrase from a ready-made catalog (a total of 59 options across 37 languages). The model runs fast and compactly, allowing it for on-device execution and integration into email clients.
Key features
- Support for 37 languages, including Russian, English, and German.
- Analysis of message context and key concepts instead of generating text from scratch.
- Compact size of about 1.2 MB thanks to storing embeddings in 8-bit integers.
- Local execution on-device to maintain privacy.
What you can do
- Integrate it into an email client to generate quick reply options for incoming emails.
- Use the train.py training script to retrain the model on your own data.
Example
python3 train.py --quick # быстрый запуск обучения и оценка на проверочных данных
python3 train.py # генерация полноценного файла replies-v1.binRead more
DetailsArchitecture: linear network with 32 hidden units and a softmax output for 59 classes.
Memory requirements: model weights are about 1.2 MB, PICAWHN1 format.
Limitations: the catalog contains a fixed number of reply options for each of the scenarios.










Comments