NightmareNet

Autonomous AI Self-Improvement through

|

Force neural networks to learn invariant structures — not memorize patterns.

WakeDreamNightmareCompress

250+

Tests Passing

4

Training Phases

7

API Endpoints

< 50ms

Avg Latency

Scroll
Try It Now

See It In Action

Watch your text transform through dream and nightmare distortions in 3 steps.

Your Text
Dream
Nightmare
How It Works

4-Phase Training Pipeline

Each cycle builds robust invariant knowledge that survives even the harshest perturbations.

01

Wake

Train on clean data to establish baseline competence and anchor model weights.

02

Dream

Soft perturbations stretch learned representations while preserving core structure.

03

Nightmare

Aggressive adversarial distortions test the model's invariant knowledge.

04

Compress

Distill robust representations into a tighter model via knowledge distillation.

Repeat each epoch → invariant structure, not surface patterns

Dream Distortion

Synonym swaps, paraphrase, soft noise

Nightmare Distortion

Adversarial attacks, char-swaps, deletions

Compression

Knowledge distillation, magnitude pruning

Get Started

3 Lines to Robustness

Add NightmareNet to your training pipeline in minutes, not days.

git clone https://github.com/Adit-Jain-srm/NightmareNet.git
cd NightmareNet
pip install -e ".[api]"
What happens each cycle (~38 min on A100)

Wake

Train on clean data

~15 min

Dream

Gentle distortions

~10 min

Nightmare

Adversarial stress test

~8 min

Compress

Prune & distill

~5 min

Interactive

Distortion Playground

Feed text through the distortion pipeline and watch it transform in real-time.

0.50
GentleAggressive

🌀 Token drops, paraphrasing — moderate restructuring

Analysis

Resilience Lab

Measure how text degrades under distortion. A resilience score of 85% means your text retains 85% of its structure under adversarial pressure.

0.20
0.70
Configuration

Training Lab

Configure training hyperparameters and preview the full phase schedule before launch.

3
3
2
1
0.25
0.80
5.0e-5
0.20

Configure & Preview

Adjust sliders and click preview to see the training schedule.

End-to-End Pipeline

Train a Hardened Model

Bring your own data → select a model → configure the sleep cycle → get a production-ready model with before/after robustness metrics.

Data Input

Upload Data

Feed your text data into the distortion pipeline. Supports .txt, .csv, and .json.

Drag & drop or click to upload

.txt, .csv, .json • Max 5MB

Architecture

Model Explorer

Inspect the transformer architecture that NightmareNet trains, distorts, and compresses.

Total Parameters124.4M

Training Capabilities

Mixed Precision

AMP + gradient checkpointing for 2× speed

Streaming Data

Process datasets larger than memory

Distributed

Multi-GPU via HuggingFace Accelerate

Multi-Arch

GPT-2, BERT, DistilBERT, and more

Default Model

GPT-2

AutoModelForCausalLM • 124M params • wikitext-2

Live

System Status

Real-time endpoint health, latency monitoring, and API diagnostics.

Checking...

Waiting...

Endpoint Health Matrix

Health

GET /api/v1/health

Dream

POST /api/v1/generate/dream

Nightmare

POST /api/v1/generate/nightmare

Robustness

POST /api/v1/evaluate/robustness

Compare

POST /api/v1/compare

Train Config

POST /api/v1/train/config

Upload

POST /api/v1/upload/text