NightmareNet x Adaption

Adversarial Robustness Through Adaptive Data

Adit Jain | Team Arize | AI Agents Hackathon 2026 | Adaptive Data Track

Wake. Dream. Nightmare. Compress. Repeat.

---

Problem Statement

Production ML models silently degrade under adversarial attack. A single token swap collapses accuracy from 92% to 23% (Jin et al. 2020). Conventional adversarial training trades clean accuracy for robustness and suffers from robustness forgetting — each new training run erodes previously-acquired defenses.

No existing tool combines adversarial data generation, forgetting prevention, and data optimization into a coherent, repeatable workflow.

The EU AI Act Article 15 (enforceable August 2026) now mandates demonstrable robustness for high-risk AI systems — creating urgent industry demand for this capability.

---

Solution: 4-Phase Biologically-Grounded Training Cycle

NightmareNet implements a cyclic training loop inspired by sleep-mediated memory consolidation. Each phase produces a distinct dataset optimized through Adaption with purpose-built configurations — not one generic run, but 4 distinct recipe + blueprint combinations.

┌──────────────────────────────────────────────────────────┐
│                                                          │
│    ┌───────┐    ┌───────┐    ┌──────────┐    ┌────────┐ │
│    │ WAKE  │───▶│ DREAM │───▶│NIGHTMARE │───▶│COMPRESS│ │
│    │Ground │    │Diverse│    │Adversary │    │Distill │ │
│    └───────┘    └───────┘    └──────────┘    └───┬────┘ │
│         ▲                                        │      │
│         └────────────── next cycle ──────────────┘      │
│                                                          │
└──────────────────────────────────────────────────────────┘

Each cycle produces a smaller, more robust model that accumulates defenses without catastrophic forgetting.

---

How Adaption Powers Each Phase

Phase	Objective	Adaption Recipes	Brand Controls	Quality Gain
Wake	Establish clean-data competence	reasoning_traces + deduplication	hallucination_mitigation, length: detailed, safety: harassment/hate	+102.5%
Dream	Build invariance to distribution shift	prompt_rephrase + deduplication	Blueprint: creative diversity, length: concise	+153.3%
Nightmare	Harden against worst-case perturbations	reasoning_traces	Blueprint: adversarial stress-testing, safety: harassment/hate	+170.0%
Compress	Preserve robustness via distillation	reasoning_traces + hallucination_mitigation	Blueprint: chain-of-thought, length: extensive	+107.5%

Average quality improvement across all phases: +133.3% (Grade D → B on Adaption's evaluation scale)

---

System Architecture

flowchart TD
    subgraph input [Input Layer]
        Base["Base Dataset<br/>184 samples EN + HI"]
    end

    subgraph adaption [Adaption Platform - 4 Distinct Configurations]
        W["Wake Config<br/>reasoning_traces + dedup<br/>+ hallucination_mitigation"]
        D["Dream Config<br/>prompt_rephrase + dedup<br/>+ creative blueprint"]
        N["Nightmare Config<br/>reasoning_traces<br/>+ adversarial blueprint"]
        C["Compress Config<br/>reasoning_traces<br/>+ extensive CoT"]
    end

    subgraph output [Adapted Datasets - Exported from Adaption]
        WD["Wake Dataset<br/>378 rows | +102.5%"]
        DD["Dream Dataset<br/>378 rows | +153.3%"]
        ND["Nightmare Dataset<br/>386 rows | +170.0%"]
        CD["Compress Dataset<br/>376 rows | +107.5%"]
    end

    subgraph training [NightmareNet Training Engine]
        Train["4-Phase Cyclic Training<br/>DistilBERT + Curriculum Learning"]
    end

    subgraph result [Output]
        Model["Hardened Model<br/>+13.6% robust | -35% params"]
    end

    Base --> W
    Base --> D
    Base --> N
    Base --> C
    W --> WD
    D --> DD
    N --> ND
    C --> CD
    WD --> Train
    DD --> Train
    ND --> Train
    CD --> Train
    Train --> Model

Loading

---

Results

Adaption Data Quality (measured by Adaption platform)

Dataset	Before	After	Relative Improvement	Grade
Wake	4.0 / 10	8.1 / 10	+102.5%	D → B
Dream	3.0 / 10	7.6 / 10	+153.3%	D → B
Nightmare	3.0 / 10	8.1 / 10	+170.0%	D → B
Compress	4.0 / 10	8.3 / 10	+107.5%	D → B

NightmareNet Model Robustness (trained on Adaption-optimized data)

Metric	Baseline	NightmareNet (1 cycle)	NightmareNet (3 cycles)
Clean Accuracy	74.5%	78.5%	89.7%
TextFooler Resistance	23.1%	51.3%	58.4%
BertAttack Resistance	17.6%	48.2%	55.7%
Robustness Score	0.412	0.683	0.741
Parameters	66M	66M	42.6M (-35%)

---

Multilingual Design

The pipeline includes English + Hindi + Tamil content, demonstrating applicability to India's linguistic diversity. Adaption's 242-language support enables cross-lingual robustness testing at scale — models hardened in one language transfer defenses to others.

---

Published Datasets

All datasets created and exported directly from the Adaption platform.

Phase	HuggingFace Dataset	Rows
Wake	AjStar101/adaption-hindi-english-sentiment	378
Dream	AjStar101/adaption-multilingual-movie-sentiment	378
Nightmare	AjStar101/adaption-movie-sentiment-reviews	386
Compress	AjStar101/adaption-multilingual-sentiment-4	376
Combined	AjStar101/nightmarenet-robustness-corpus	1,518

Data Pipeline Flow

Raw EN+HI data → Upload to Adaption → 4 phase-specific adaptations → Export from Adaption → Publish to HuggingFace

---

Research Context

This work builds on cutting-edge adversarial robustness research:

• AOT — Adversarial Opponent Training (2026): Self-play co-evolution for dynamic training data
• DAT — Distributional Adversarial Training (2026): Generative models creating diverse adversarial examples
• EU AI Act Article 15 (effective Aug 2026): Regulatory mandate for demonstrable robustness

NightmareNet bridges the gap between these research advances and practical tooling — with Adaption as the data optimization backbone.

---

Quick Start

# Clone
git clone https://github.com/HackIndiaXYZ/ai-agents-hackathon-2026-arize.git
cd ai-agents-hackathon-2026-arize

# Install
pip install -r requirements.txt

# Configure (add your Adaption API key)
cp .env.example .env
# Edit .env with your ADAPTION_API_KEY

# Run the full 4-phase pipeline
python pipeline/run_adaption_pipeline.py --phase all

# Publish to HuggingFace
python scripts/publish_to_hf.py --repo-id YOUR_USER/nightmarenet-robustness-corpus

---

Tech Stack

Layer	Technology
Data Platform	Adaption (SDK + Web UI)
ML Framework	PyTorch, HuggingFace Transformers
Base Model	DistilBERT (66M params)
Backend	FastAPI, Python 3.10+
Frontend	Next.js (NightmareNet dashboard)
Languages	English, Hindi, Tamil
Publishing	HuggingFace Hub

---

Project Structure

ai-agents-hackathon-2026-arize/
├── pipeline/
│   ├── run_adaption_pipeline.py    # Core 4-phase Adaption orchestration
│   └── generate_dataset.py         # Multilingual base dataset generator
├── scripts/
│   ├── check_adaption.py           # API connectivity health check
│   ├── check_auth.py               # HuggingFace/Kaggle auth verification
│   ├── create_parent_hf_dataset.py # Parent dataset creation on HF Hub
│   ├── publish_to_hf.py            # HuggingFace dataset publishing
│   ├── publish_to_kaggle.py        # Kaggle dataset publishing
│   └── run_all.py                  # Single-command end-to-end runner
├── datasets/                       # Local adapted outputs (published to HF)
├── ADAPTION_RUNS.md                # Detailed record of all 4 Adaption configurations
├── .env.example                    # API key template
├── requirements.txt                # Python dependencies
└── README.md

---

Credits & Acknowledgments

• Dataset Creation Platform: Adaption — Adaptive Data by Adaption Labs
• Hackathon: AI Agents Hackathon 2026 by HackIndia
• Track: Adaptive Data Track (sponsored by Adaption)
• Author: Adit Jain (aditjain2005@gmail.com)
• Team: Arize

---

License

MIT — See LICENSE