Software characteristics

Roadmap

Calendar

  • Experimental generative for April.
  • Stable classical version planned for mid-June (+ Docker).

Next Steps

  • Multilabel workflow + bert-fine tuning
  • Create Python wrapper
  • Write documentation + tutorial
  • Optimize vizualisation for large dataset ⚙
  • Create a easy/medium/pro mode
    • need definition
  • Define the monitor panel
    • need definition
  • Optimize GPU management (prediction)

Enhancements

  • Build Docker image ⚙
  • Integrate genAI tools ⚙
  • Add carbon count
  • Add new models (Modernbert, Phi3, pleias, ...)
  • Refactor design (better ergonomy / colors / etc.) ⚙
  • Better data set management (expand)
  • Animate community on Discord

Possibilities

  • Attribute specific task to users

Architecture

This is a collection of technical points/choices for the app

Overall architecture :

  • backend : Python/FastAPI
  • frontend : React/Typescript

Backend

  • config.yaml define the parameters at the server launch
  • The unit is the project, composes of different classes
    • Features
    • Schemes
    • Simplemodels
    • Bertmodels
    • Users
  • CPU/GPU bound computation is managed in separated processes with a queue
  • State of the service is checked at each request (with a threshold)

Data management

  • Tabular data is stored as separated parquet files divided in train / test / complete
  • SQLite database to manage annotations/parameters/users/logs
  • Projects are loaded into memory to facilitate computation (filter, etc.)
    • Unloaded after one day
  • Bert models are saved in dedicated filesystems

Processes

  • ProcessPoolExecutor with workers
    • https://superfastpython.com/processpoolexecutor-in-python/
  • Different type of parallel process : training ; predicting
  • Only one process possible by user/project

Users role

  • Role-Based Access Control (RBAC) - 3 roles : root, manager, annotator
  • Authentification with OAuth2 and token in header
    • Table of valid tokens
  • A table of authorization defines the relation users/projects
  • Different uses can modify a same project : no lock

Select element to annotate

The selection combines different strategy : filters and/or active learning.

Active learning is a prediction with a model trained on already annotated data.

  • Different modes of selection

    • deterministic
    • aleatory
    • maxprob for a label
    • max entropy

  • Pipeline of choice

    • sample (tagged, untagged, all)
    • regex
    • proba / entropie

Frontend

State management

  • Each project is described by its general state (not user specific)
    • Computed/computing elements