Description
ej is a decision model that fits into a compact 11 MB file, designed to handle routing, yes/no, and priority questions about text. It operates efficiently on a CPU, eliminating the need for an API or GPU, which makes it ideal for offline use. The model provides probability-based answers for each option, allowing users to set thresholds based on their specific needs.
Read More
The model is fine-tuned from the intfloat/e5-small-v2 and trained on 9,719 records, including Typed Decisions workflows, Banking77, CLINC150, and GoEmotions datasets. It is particularly strong in workflows it was trained on, achieving an accuracy of .721, and performs well on support tickets in new writing styles with an accuracy of .683.
ej is designed to be safe and efficient, with SHA-256 checks ensuring the integrity of the model file. It operates at a speed of approximately 214 ms per record on a single CPU thread, with peak memory usage around 558 MB. Users can also train their own models using the provided tools, allowing for customization and adaptation to specific tasks.
The model is licensed under Apache-2.0 for the code and CC BY-SA 4.0 for the weights, ensuring open access and the ability to share modifications under the same license. ej is suitable for teams that need to make repetitive decisions about text, especially when operating on their own servers or offline.
- CategoryAI Models & LLMs
- Pricing ModelOpen source
- DeploymentSelf-hosted
- PlatformsWindows, macOS, Linux
ej's Core Features
Compact 11 MB model file
Runs on CPU, offline
Probability-based decision making
Fine-tuned from intfloat/e5-small-v2
Trained on diverse datasets
Open weights and code
Apache-2.0 and CC BY-SA 4.0 licenses
SHA-256 checked model file
How to use ej?
Install: pip install ejai
Load: Initialize the model on your CPU
Query: Pass text queries for decision making
Adapt: Train with your own data if needed
ej's Use Cases
- Text Routing
- Priority Setting
- Yes/No Decisions
- Offline Processing
- Custom Model Training
