arc-training 4.0.0

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

ARC - Automatic Recovery Controller

Auto-detect and recover from neural network training failures.

ARC automatically detects NaN/Inf losses, gradient explosions, OOM errors, and silent failures—then recovers your training without losing progress.

Key Results

Metric	ARC v4.0
Recovery Rate	100% (on test suite)
Overhead	27% (ARC Lite)
vs torchft	3/3 vs 1/3 recoveries
Max Model Size	1.5B params

🚀 Quick Start

Installation

pip install arc-training

Basic Usage (3 lines!)

from arc import WeightRollback

# Initialize
arc = WeightRollback(model, optimizer)

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

    # ARC handles everything
    action = arc.step(loss)

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

Lightning Integration

from arc.integrations import ARCCallback

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

📊 Configurations

Config	Overhead	Use Case
ARC Lite	27%	Production training
ARC Full	44%	Debugging unstable runs

🛡️ What ARC Handles

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

📈 Benchmarks

Recovery Rate: 100% (160/160 induced failures)
Statistical Significance: p < 0.001
Models Tested: CNN, ViT, Transformer, Diffusion (up to 1.5B params)

🔗 Links

• Documentation
• Paper
• Efficiency Report

📜 Citation

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

📄 License

AGPL-3.0 License - see LICENSE for details.

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