dsf-aml-sdk 2.2.0

DSF AML SDK — Automated ML Robustness & Failure Correction Framework

DSF AML SDK

Automated ML Robustness & Training Data Generation

Generate critical training variants from production failures and edge cases. Accelerate model retraining with automatically curated datasets.

---

🎯 Primary Use Cases

1. Production Failure Recovery

Challenge: ML/LLM models fail on edge cases. Manual correction is slow.

Solution: Generate critical variants from failures for rapid retraining.

from dsf_aml_sdk import AMLSDK

sdk = AMLSDK(license_key='your_key', tier='professional')

# Production failure detected
failed_case = {'metric_a': 0.60, 'metric_b': 500, 'metric_c': 0.20}

# Generate variants
variants = sdk.generate_variants(
    seed=failed_case,
    config=your_config,
    count=20
)

# Use variants['samples'] for retraining

Output: Labeled data points similar to failure case for model robustness improvement.

---

2. Preventive Dataset Curation

Challenge: Models trained on clean data fail on boundary cases.

Solution: Pre-generate datasets focused on decision boundaries.

# Identify high-impact regions
seeds = sdk.identify_high_impact_regions(
    dataset=training_data,
    config=config,
    focus_percent=0.1
)

# Generate boundary variants
boundary_data = sdk.generate_boundary_variants(
    config=config,
    source_data=training_data,
    variants_per_seed=10
)

# Train with boundary_data

---

3. Training Data Generation

Challenge: Creating labeled datasets is expensive.

Solution: Generate synthetic labeled datasets at scale.

# Generate labeled samples
result = sdk.generate_training_data(config, samples=1000)

# Export (Enterprise)
dataset = sdk.export_dataset()

# Train your models with generated data

---

📦 Installation

pip install dsf-aml-sdk

---

🧩 Quick Start

from dsf_aml_sdk import AMLSDK

sdk = AMLSDK(license_key='your_key', tier='professional')

# Define evaluation config
config = {
    'metric_a': {
        'reference_value': 0.95,
        'params': {
            'importance': 2.5,
            'sensitivity': 2.0
        }
    },
    'metric_b': {
        'reference_value': 100,
        'params': {
            'importance': 1.8,
            'sensitivity': 1.5
        }
    }
}

# Report failure
failed_input = {'metric_a': 0.60, 'metric_b': 500}

# Generate corrections
fix = sdk.generate_variants(failed_input, config, count=20)

print(f"Generated {len(fix['samples'])} variants")

---

📊 Execution Metrics

Operations return performance metrics:

{
  "tier": "professional",
  "evaluations": 62,
  "threshold": 0.6698,
  "persistence": "active",
  "statistics": {
    "avg": 0.7296,
    "min": 0.5217,
    "max": 0.8467
  }
}

---

🆚 Tier Comparison

Feature	Community	Professional	Enterprise
Variant Generation	Limited	✅	✅
Preventive Datasets	Limited	✅	✅
Batch Operations	❌	✅ (≤1000)	✅ (≤1000)
Data Export	❌	✅	✅
Full Pipeline	❌	❌	✅

---

📖 Core Methods

Variant Generation

sdk.generate_variants(seed: dict, config, count=20) → dict

Returns:

{
  "status": "completed",
  "total": 20,
  "samples": [...],
  "metrics": {...}
}

High-Impact Region Identification

sdk.identify_high_impact_regions(dataset, config, focus_percent=0.1) → dict
sdk.generate_boundary_variants(config, source_data, **kwargs) → dict

Training Data Generation

sdk.generate_training_data(config, samples=1000) → dict
sdk.export_dataset() → dict  # Enterprise only

Evaluation

# Single evaluation
result = sdk.evaluate(data, config)

# Batch evaluation
results = sdk.batch_evaluate(data_points, config)

---

🔧 Configuration Structure

config = {
    "feature_name": {
        "reference_value": <target_value>,
        "params": {
            "importance": <float>,    # Feature weight
            "sensitivity": <float>    # Deviation tolerance
        }
    }
}

Example Configuration

config = {
    'metric_primary': {
        'reference_value': 650,
        'params': {
            'importance': 2.5,
            'sensitivity': 2.0
        }
    },
    'metric_secondary': {
        'reference_value': 60000,
        'params': {
            'importance': 2.0,
            'sensitivity': 1.8
        }
    }
}

---

🛠️ Complete Workflow

import pandas as pd
from dsf_aml_sdk import AMLSDK

# Initialize
sdk = AMLSDK(license_key='your_key', tier='professional')

# Load data
df = pd.read_csv('data.csv')
data = df[['metric_a', 'metric_b', 'metric_c']].head(100).to_dict('records')

# Config
config = {
    'metric_a': {
        'reference_value': 650,
        'params': {'importance': 2.5, 'sensitivity': 2.0}
    },
    'metric_b': {
        'reference_value': 60000,
        'params': {'importance': 2.0, 'sensitivity': 1.8}
    }
}

# 1. Fix failure
failed = {'metric_a': 580, 'metric_b': 35000}
fix = sdk.generate_variants(seed=failed, config=config, count=10)

# 2. Preventive dataset
seeds = sdk.identify_high_impact_regions(data[:50], config, focus_percent=0.15)
boundary = sdk.generate_boundary_variants(config, data[:50], variants_per_seed=20)

# 3. Generate training data
result = sdk.generate_training_data(config, samples=200)

# 4. Evaluate
test = data[0]
result = sdk.evaluate(test, config)

---

⚠️ Important Notes

Client Responsibility:
Clients must validate model performance and compliance with applicable regulations. This SDK is a data generation tool and does not make autonomous decisions.

Data Processing:
All generation logic executes server-side. SDK exposes configuration interface only.

Generated Data:
Synthetic data is based on client-provided configurations and source datasets. Clients control all inputs and validation.

---

📞 Support

Licensing: [e-mail]
Technical Docs: Available under NDA
Enterprise: [e-mail]

---

📋 Credits

Technology Architect: Jaime Alexander Jimenez

---

© 2025 DSF UpTech. All rights reserved.