arc-training 4.0.1

Automatic Recovery Controller - Auto-detect and recover from neural network training failures

ARC - Autonomous Recovery Controller

A real-time fault-tolerance framework for neural network training.

ARC monitors training signals — gradients, loss curvature, Fisher Information — to predict and recover from failures before they crash your run. It uses a Mamba-based state-space model with Evidential Deep Learning for uncertainty-aware failure prediction.

Key Results

Metric	ARC v4.0
Recovery Rate	100% on core failure types (NaN, Inf, explosion)
Overhead	~35% (small models), higher on larger models
vs torchft	3/3 vs 1/3 recoveries
Models Tested	YOLOv11, DINOv2, Llama-Style, SD-UNet (up to 33M params)

Quick Start

Installation

pip install arc-training

Basic Usage

from arc import Arc

# Wrap your model and optimizer
controller = Arc(model, optimizer)

# Training loop
for batch in dataloader:
    loss = model(batch)

    # ARC monitors, predicts, and recovers automatically
    action = controller.step(loss)

    if not action.rolled_back:
        loss.backward()
        optimizer.step()

Lightning Integration

from arc import ArcCallback

trainer = pl.Trainer(
    callbacks=[ArcCallback()]
)

Configurations

Config	Overhead	Use Case
ARC Lite	Lower	Production training
ARC Full	Higher	Debugging unstable runs

Failure Coverage

Failure Type	Detection	Recovery	Status
NaN/Inf Loss	Yes	Yes	Validated
Loss Explosion	Yes	Yes	Validated
Gradient Explosion	Yes	Yes	Validated
OOM (all stages)	Yes	Yes	Validated
Accuracy Collapse	Yes	Partial	Detection only
Mode Collapse	Yes	Partial	Detection only

Benchmarks

Core Failure Recovery: 100% (NaN, Inf, explosion across all tests)
Modern Model Recovery: 5/8 induced failures recovered
torchft Comparison:    ARC 3/3 vs torchft 1/3
Models Tested:         YOLOv11, DINOv2-Small, Llama-Style, SD-UNet

Links

• Efficiency Report

Citation

@software{arc2026,
  title={ARC: Autonomous Recovery Controller for Neural Network Training},
  author={Kaushik, Aryan},
  year={2026},
  url={https://github.com/a-kaushik2209/ARC}
}

License

AGPL-3.0 License - see LICENSE for details.

Copyright (c) 2026 Aryan Kaushik. All rights reserved.