The_Recovery_From_WithIn_10k

Creator: WithIn Us Ai (WithinUsAI)
Version: 1.0
Size: 10,000 unique professional examples
Format: JSONL
License: MIT

Description

The_Recovery_From_WithIn_10k is a frontier-quality dataset designed to train advanced autonomous agentic language models on professional failure diagnosis, root cause analysis, recovery planning, fallback strategies, and verification after incidents.

Every example teaches the model how to detect failures, perform rigorous root cause analysis, design safe and effective recovery plans, identify fallback options, and verify successful recovery — all while minimizing further damage and preventing recurrence.

Purpose

Train LLMs to master:

• Failure detection and diagnosis
• Root cause analysis under pressure
• Safe, prioritized recovery planning
• Fallback and contingency design
• Post-recovery verification and prevention of recurrence

Training Benefits

This dataset improves:

• Robustness and error recovery capability
• Diagnostic reasoning and root cause analysis
• Safe decision making during incidents
• Long-horizon reliability and resilience
• Autonomous task completion even after setbacks
• Professional incident response and recovery skills

Dataset Structure

JSONL file where each line is a JSON object with the following schema:

{
  "instruction": "string - The recovery or incident response request",
  "input": "string - Observed symptoms, logs, error messages, and current state",
  "output": {
    "thinking": "<think>Detailed professional reasoning trace demonstrating diagnosis, root cause analysis, prioritization, planning, and decision making</think>",
    "failure_detected": "string - Clear description of the detected failure and its impact",
    "root_cause_analysis": ["string", ...] - Key findings from root cause investigation,
    "recovery_plan": ["string", ...] - Ordered, safe recovery steps,
    "fallback_options": ["string", ...] - Alternative recovery paths if primary plan fails,
    "verification_steps": ["string", ...] - Steps to confirm successful recovery and prevent recurrence
  }
}

Schema Documentation

• instruction: Request to recover from a specific failure or incident.
• input: Symptoms, logs, errors, and system state provided for diagnosis.
• output.thinking: Professional reasoning inside tags showing how the system diagnosed the issue, performed root cause analysis, prioritized actions, designed the recovery plan, and decided on verification.
• output.failure_detected: Precise description of what failed and the business/operational impact.
• output.root_cause_analysis: List of root causes and contributing factors identified.
• output.recovery_plan: Ordered list of actionable recovery steps designed to restore service safely.
• output.fallback_options: Alternative paths if the primary recovery plan encounters issues.
• output.verification_steps: Concrete steps to confirm the system is healthy and the root cause has been addressed.

Example Records

Example 1: ML Training Failure Recovery

{
  "instruction": "Recover from a critical failure during weekly retraining of the credit risk prediction model.",
  "input": "Training job crashed after 4 hours with OOM error. Last successful checkpoint from 3 hours ago. New transaction data for the week has been ingested. Production model is still serving but will degrade if not updated soon.",
  "output": {
    "thinking": "<think>The failure is an OOM during training on the new weekly data. Root cause is likely increased data volume or feature complexity this week combined with insufficient memory allocation. Primary recovery: resume from last checkpoint with increased memory or reduced batch size. Fallback: use previous week's model with monitoring for drift. Verification must confirm training completes successfully and new model passes all validation gates before promotion.</think>",
    "failure_detected": "OOM crash during weekly model retraining, blocking the update of the production credit risk model.",
    "root_cause_analysis": [
      "Memory allocation insufficient for this week's data volume and feature set",
      "No automatic memory scaling or early stopping on resource exhaustion",
      "Checkpointing was enabled but recovery path not tested in this pipeline"
    ],
    "recovery_plan": [
      "Resume training from last successful checkpoint with increased instance memory or reduced batch size",
      "Monitor resource usage and set alerts for future runs",
      "Validate new model on holdout set and drift detection before promotion",
      "Update pipeline to request larger instance type or implement gradient checkpointing for future runs",
      "Document the incident and add automated recovery logic to the training orchestration"
    ],
    "fallback_options": [
      "Promote previous week's model with tightened monitoring and drift alerts",
      "Manually trigger training on a larger instance type",
      "Temporarily increase inference compute while retraining completes"
    ],
    "verification_steps": [
      "Confirm training job completes successfully with new parameters",
      "Validate model performance on recent holdout data meets or exceeds baseline",
      "Run full drift detection and compliance checks on the new model",
      "Monitor production metrics for 24 hours after promotion to confirm stability"
    ]
  }
}

(Additional 9,999 unique professional examples follow the same rigorous standard across all recovery domains.)

Domain Coverage

Software bugs, model training failures, deployment failures, infrastructure outages, API failures, database corruption, data quality issues, research dead ends, business process failures, operational disruptions, cloud incidents, Linux troubleshooting, cybersecurity defense incidents, project delays, workflow breakdowns.

Quality Standards

• Exactly 10,000 unique examples
• No duplicates, near-duplicates, placeholders, templates, or low-effort content
• Every example is realistic and provides genuine frontier-level training signal for recovery agents
• Thinking traces demonstrate professional senior-level diagnostic and recovery reasoning
• Recovery plans are safe, prioritized, and include verification
• Suitable for training state-of-the-art autonomous agentic models that recover gracefully from failures

Deduplication Process

All examples were generated with aggressive semantic and structural deduplication. No two examples share the same failure signature, root cause structure, or recovery sequence beyond necessary professional terminology.

Validation Process

Every example was validated for:

• Logical consistency between thinking, failure description, root cause analysis, recovery plan, fallbacks, and verification
• Professional tone and incident-response-ready structure
• Coverage of target failure domains
• Absence of any forbidden content types

Recommended Training Usage

Use for supervised fine-tuning of agentic models on failure recovery and resilience. Preserve the full <think> reasoning for Chain-of-Thought training. Ideal for building robust agents that diagnose, recover, and verify their own work after setbacks.

Pair with planning and verification datasets for complete autonomous agent training (plan → execute → verify → recover if needed).

Citation

@misc{withinusai2026recoveryfromwithin,
  title={The_Recovery_From_WithIn_10k: Frontier Professional Failure Recovery Dataset for Agentic LLMs},
  author={WithIn Us Ai},
  year={2026},
  howpublished={\url{https://huggingface.co/datasets/WithinUsAI/The_Recovery_From_WithIn_10k}}
}

License

MIT License. Free for research and commercial use with attribution.

Acknowledgements

Created by WithIn Us Ai (WithinUsAI) with the goal of advancing robust, self-recovering autonomous agentic systems.

---

WithinUsAI — Building the highest quality datasets for frontier agentic intelligence.