🚨 Llama-3.1-8B-SREBench-GRPO
This model is a fine-tuned version of Llama 3.1 8B, aligned specifically for Autonomous Site Reliability Engineering (SRE). It was developed by Krishna Singh (SLRTCE) for the Meta PyTorch OpenEnv Grand Finale.
🏗️ Model Description
Unlike generic LLMs that "symptom chase," this model is trained to perform causal reasoning across a 6-service microservice architecture. It is designed to work within the SREBench environment, an OpenEnv-compliant world model featuring stochastic telemetry and "Victim Logging".
🧠 Training Procedure: GRPO RL
The model was aligned using Generative Reward-Optimized (GRPO) training on an NVIDIA A100 GPU. We utilized a 3-part reward signal to defeat "Reward Hacking":
- Investigation Reward (+0.15): Bonus for checking logs/metrics before acting.
- Anti-Shotgun Penalty (-0.20): Heavy penalty for restarting services without diagnostic evidence.
- Format Reward (+0.20): Ensuring strict adherence to JSON action schemas.
🚀 Usage Instructions
This model is designed to output valid JSON actions for the OpenEnv /step endpoint.
Input Format
The model expects a system prompt defining its SRE capabilities and a user prompt containing the current system_dashboard and alert_message.
Example Output
1{
2 "action_type": "remediate",
3 "command": "scale_up",
4 "target": "database-replica",
5 "params": {}
6}
📊 Performance
- Success Rate (Compound Incidents): 92%
- Reward Improvement: +140% over heuristic baselines
⚖️ Evaluation
To evaluate this model against the SREBench environment, use the official orchestrator:
python run_multi_agent_eval.py --model your-username/Llama-3.1-8B-SREBench-GRPO
Developed by Krishna Singh | Solo Competitor | Meta PyTorch OpenEnv Hackathon 2026
