Coflect

Agentic Deep Learning Framework

HITL · Human In Loop Training

---

Coflect is a human-in-the-loop (HITL) training framework built for fast training loops.

Core principles:

• Trainer hot path stays lightweight (small JSON events only).
• Heavy XAI runs in a separate worker process.
• Visualization and explainability should not block training.

• More About Coflect
  • Non-blocking HITL Architecture
  • Live Feedback During Training
  • Torch-First, Multi-Backend Roadmap
  • Fast and Lean by Design
• Installation
  • PyPI
  • From Source (Contributor Setup)
• Getting Started
  • Notebook Workflows (Primary)
  • Quick Demo (CLI)
• Resources
• Releases and Contributing
• License

More About Coflect

Non-blocking HITL Architecture

Coflect splits responsibilities across processes:

• backend: event ingest, WebSocket broadcast, feedback storage, XAI queue
• trainer: training loop + lightweight telemetry/XAI requests
• xai worker: attribution rendering outside trainer hot path
• forecast worker: CPU-only top-k likely failure ranking
• UI: live metrics, overlays, pause/resume, ROI and text feedback

Live Feedback During Training

Coflect supports live human intervention with:

• text instruction parsing (deterministic parser)
• ROI-guided focus updates
• pause/resume control from the UI
• periodic and mistake-focused XAI request paths

Torch-First, Multi-Backend Roadmap

Current release line:

• Torch: primary production path
• TensorFlow/Keras: MVP runtime path
• JAX: scaffolded path

Fast and Lean by Design

• Trainer emits compact events only.
• Heavy operations (XAI render/forecast logic) are offloaded.
• Snapshot sync and XAI requests are asynchronous.

Installation

PyPI

pip install --upgrade coflect

Optional framework extras:

pip install "coflect[tensorflow]"
pip install "coflect[jax]"

Python requirement: >=3.10.

From Source (Contributor Setup)

git clone https://github.com/Coflect/Coflect.git
cd Coflect
python -m venv .venv
source .venv/bin/activate
pip install -e .
# optional contributor tooling:
# pip install -e .[dev]
# optional framework extras:
# pip install -e .[tensorflow]
# pip install -e .[jax]

Getting Started

Notebook Workflows (Primary)

Use these notebooks for daily deep learning experiments:

• Torch: examples/hitl/01_hitl_module_quickstart.ipynb
• TensorFlow/Keras: examples/hitl/02_hitl_tensorflow_keras_workflow.ipynb

Both notebooks include bridge-based UI integration, feedback handling, and Live/XAI wiring.

Quick Demo (CLI)

For a fast end-to-end demo:

Torch quickstart (CIFAR-10 cat-vs-dog only):

coflect-hitl-run \
  --backend torch \
  --dataset cifar10_catsdogs \
  --data_root ./data \
  --download_data \
  --steps 5000 \
  --xai_every 100 \
  --forecast_every 20

Dataset note:

• class 3 (cat) -> label 0
• class 5 (dog) -> label 1

TensorFlow/Keras MVP path:

coflect-hitl-run \
  --backend tensorflow \
  --steps 5000 \
  --xai_every 100 \
  --forecast_every 20

Open UI at http://localhost:8000.

Resources

• Architecture: docs/ARCHITECTURE.md
• Support policy: SUPPORT_MATRIX.md
• Example index: examples/README.md
• HITL examples: examples/hitl/README.md
• Release playbook: docs/RELEASE_PYPI.md
• Launch checklist: docs/LAUNCH_CHECKLIST.md

Releases and Contributing

• Releases: https://github.com/Coflect/Coflect/releases
• Contributing guide: CONTRIBUTING.md
• Security policy: SECURITY.md
• Code of Conduct: CODE_OF_CONDUCT.md

License

Apache-2.0 (LICENSE).