miyad316/sre-agentic-trajectories-devops

SRE & DevOps Agentic Trajectory Dataset (2,000-Row Preview)

Notice: This repository provides an open 2,000-row preview sample licensed under Creative Commons Attribution-NonCommercial 4.0 (CC-BY-NC-4.0) for academic evaluation and research.
🛒 For Commercial Model Training & Full 5,000 Production Dataset ($49 $9.99 Launch Special):
👉 Click Here to Purchase on Gumroad ($9.99 Instant Download) (Immediate delivery of full dataset + Commercial License).

---

🚀 Overview

Training small language models (3B to 8B) to reliably operate as autonomous infrastructure agents requires complex, multi-turn tool-use data. Most existing open datasets provide single-turn happy-path queries on trivial tools (weather, calculators).

This dataset provides enterprise-grade incident remediation trajectories across real-world DevOps environments:

• Kubernetes (kubectl): OOMKilled crashes, CrashLoopBackOff, readiness probe timeouts, pod resource patching.
• Linux OS: Inode/disk exhaustion, zombie processes, socket state leaks (TIME_WAIT).
• APM & Cloud Metrics: Datadog/CloudWatch p99 latency queries, 5xx error spikes, dependency tracing.
• PostgreSQL: Table lock contentions, connection pool exhaustion, transaction cancellation.

🛡️ Key Features

• 26% Self-Healing & Error-Recovery Turns: Agents handle 404s, 504 timeouts, and rate limits gracefully, re-orienting cluster state before executing remediation.
• 100% Strict Schema Validation: Every tool call adheres strictly to Draft-7 JSON schemas.
• Two-Tier Quality Filter: Cleaned via deterministic schema checks and LLM-as-a-Judge validation.
• Zero Quantization Noise: Generated on AMD Instinct MI300X (192GB VRAM) running native BF16 Qwen/Qwen2.5-72B-Instruct.

---

📦 How to Load in Python

from datasets import load_dataset

# Load preview sample directly
dataset = load_dataset("json", data_files="sample_2000.jsonl")

# Inspect first incident
sample = dataset["train"][0]
print("Incident ID:", sample["id"])
print("Category:", sample["metadata"]["category"])
print("Turns:", len(sample["messages"]))

---

📊 Benchmark Fine-Tuning Uplift

Fine-tuning Qwen-2.5-7B on this SRE trajectory dataset yields significant improvements on multi-turn function calling:

• Berkeley Function Calling Leaderboard (BFCL): +16.4% Executable Accuracy (58.2% → 74.6%).
• Tool Failure Recovery Rate: +58.8% Success on unexpected API 5xx/404 (22.4% → 81.2%).
• Hallucinated Tool Calls: Reduced by -12.7% (14.8% → 2.1%).

---

💼 Full Commercial Dataset ($49 $9.99 Launch Special)

Need the un-sampled 5,000 production trajectories (4,938 strictly validated) with commercial training rights?

What is Included in the Full Package:

1. Full 5,000 Production Trajectories (.jsonl): Both OpenAI Function Calling format and ShareGPT / LLaMA-Factory format.
2. LLaMA-Factory & Axolotl Configs: Ready-to-train dataset_info.json.
3. Full Commercial Rights: Perpetual worldwide license to train proprietary models, fine-tune SLMs, and monetize model weights.
4. Verified QA Audit Report: Provenance, schema audit, and rejection documentation.

💳 Instant Purchase ($49 $9.99 Instant Download):

You can purchase and immediately download the complete package via Gumroad:

👉 Buy Commercial License & Full Dataset on Gumroad ($9.99)

• Accepts Credit / Debit Cards, Apple Pay, Google Pay, and PayPal.
• Instant digital delivery of the 7.4 MB package upon checkout.
• For enterprise invoicing or custom trajectories, email miyadislam316@gmail.com.