hs-zz27/incident-commander-1p5b-v5

Incident Commander 1.5B v5

A fine-tuned Qwen2.5-1.5B-Instruct model trained to act as an AI SRE (Site Reliability Engineer) incident commander for production outage resolution.

Training Pipeline

1. SFT Warm-start — Supervised fine-tuning on expert incident response trajectories
2. GRPO — Group Relative Policy Optimization with shaped rewards for diagnosis-before-action behavior

Files

Directory	Description	Size
sft_merged_1p5b_v5/	SFT-merged base model (full weights)	~3 GB
trained_model_1p5b_v5/	GRPO LoRA adapter (final checkpoint)	~160 MB

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

base = AutoModelForCausalLM.from_pretrained(
    "hs-zz27/incident-commander-1p5b-v5",
    subfolder="sft_merged_1p5b_v5",
)
tokenizer = AutoTokenizer.from_pretrained(
    "hs-zz27/incident-commander-1p5b-v5",
    subfolder="sft_merged_1p5b_v5",
)
model = PeftModel.from_pretrained(
    base,
    "hs-zz27/incident-commander-1p5b-v5",
    subfolder="trained_model_1p5b_v5",
)

Environment

Trained and evaluated on the Incident Commander OpenEnv — an RL environment simulating cascading production outages across microservices.

Tasks

• Single service failure (easy)
• Cascading failure (medium)
• Hidden root cause (hard)
• Chaos cascade (hard)
• Multi-root cause (expert)