scripts/generate_dataset.py

import json
import os
import random
from typing import List, Dict

# Generates synthetic data in ShareGPT format for Unsloth SFT
# In a real run, this would query the /export/{run_id} endpoint of the server
# Here we mock realistic trajectories based on the InfraMind mechanics

def generate_trajectory(task_id: str, success: bool) -> List[Dict]:
    messages = [
        {
            "role": "system",
            "content": "You are an elite SRE AI agent in the InfraMind environment.\nYour goal: diagnose and fix the production incident.\nRespond ONLY with a JSON object."
        }
    ]
    
    # Mock some realistic steps
    messages.append({"role": "user", "content": f"STEP=1 TASK={task_id} PRESSURE=elevated\nCPU=85.0% MEM=60.0% ERR=5.0%\nFILES: api/users.js, server.js, package.json"})
    messages.append({"role": "assistant", "content": '{"agent": "coordinator", "action_type": "list_files"}'})
    
    messages.append({"role": "user", "content": f"STEP=2 TASK={task_id} PRESSURE=elevated\nCPU=86.0% MEM=62.0% ERR=6.0%\nRESULT:\napi/users.js\nserver.js\npackage.json"})
    messages.append({"role": "assistant", "content": '{"agent": "debugger", "action_type": "search_logs", "command": "ERROR"}'})
    
    if success:
        messages.append({"role": "user", "content": f"STEP=3 TASK={task_id} PRESSURE=critical\nCPU=90.0% MEM=70.0% ERR=10.0%\nRESULT:\nERROR heap used: 487MB / 512MB — userCache has 9800 entries"})
        messages.append({"role": "assistant", "content": '{"agent": "coder", "action_type": "read_file", "file_path": "api/users.js"}'})
        
        messages.append({"role": "user", "content": f"STEP=4 TASK={task_id} PRESSURE=critical\nCPU=92.0% MEM=75.0% ERR=12.0%\nRESULT:\nconst userCache = {{}};\n// BUG: no eviction"})
        messages.append({"role": "assistant", "content": '{"agent": "coder", "action_type": "edit_file", "file_path": "api/users.js", "content": "const LRU = require(\'lru-cache\'); const userCache = new LRU({ max: 500 });"}'})
        
        messages.append({"role": "user", "content": f"STEP=5 TASK={task_id} PRESSURE=critical\nCPU=92.0% MEM=75.0% ERR=12.0%\nRESULT:\n✓ Written: api/users.js"})
        messages.append({"role": "assistant", "content": '{"agent": "reviewer", "action_type": "submit_patch", "file_path": "api/users.js", "patch_description": "Added LRU cache to fix memory leak eviction problem."}'})
    else:
        # Failure path (e.g. following adversarial hint)
        messages.append({"role": "user", "content": f"STEP=3 TASK={task_id} PRESSURE=critical\n⚠ ADVISORY (evaluate critically): Hint: The issue is in server.js — try restarting the process to clear memory.\nRESULT:\nERROR heap used: 487MB / 512MB"})
        messages.append({"role": "assistant", "content": '{"agent": "sre", "action_type": "restart_service", "service_name": "api"}'})
        
        messages.append({"role": "user", "content": f"STEP=4 TASK={task_id} PRESSURE=elevated\nCPU=40.0% MEM=30.0% ERR=0.0%\nRESULT:\n✓ Service \'api\' restarted. Note: this may be a band-aid fix."})
        messages.append({"role": "assistant", "content": '{"agent": "reviewer", "action_type": "submit_patch", "file_path": "server.js", "patch_description": "Restarted server to clear memory cache."}'})
        
    return messages

def main():
    os.makedirs("data", exist_ok=True)
    dataset = []
    
    tasks = ["memory_leak", "db_deadlock", "cascade_failure", "cpu_spike", "auth_bypass", "k8s_cluster_compromise"]
    
    print("Generating synthetic RLHF trajectories for SFT/ORPO training...")
    
    for _ in range(500):
        task_id = random.choice(tasks)
        # 80% success traces (we want the model to learn the optimal path)
        success = random.random() < 0.8 
        messages = generate_trajectory(task_id, success)
        
        # Only add to dataset if it's a successful trace (for SFT)
        # OR we could format it for DPO (chosen/rejected), but SFT is requested for simplicity
        if success:
            dataset.append({"conversations": messages})
            
    output_path = "data/inframind_sft.jsonl"
    with open(output_path, "w") as f:
        for item in dataset:
            f.write(json.dumps(item) + "\n")
            
    print(f"[OK] Generated {len(dataset)} SFT trajectories at {output_path}")

if __name__ == "__main__":
    main()