Lebowski17/devops-lora-mistral

DevOps Troubleshooting Assistant (LoRA)

A specialized LoRA fine-tuned adapter built on top of Mistral-7B-Instruct-v0.2, optimized for diagnosing and resolving infrastructure, deployment, and development bottlenecks.

This model acts as a secondary pair of eyes for developers, helping debug:

• Containerization: Docker exit codes, volume mounting, and networking issues.
• Version Control: Git merge conflicts, detached HEAD states, and workflow errors.
• System Admin: Linux permission hurdles, shell scripting bugs, and cron job failures.
• Cloud & CI/CD: Deployment pipeline failures and environment variable misconfigurations.
• Backend: API bottlenecks and database connection timeouts.

---

🛠 Training Details

• Base Model: mistralai/Mistral-7B-Instruct-v0.2
• Method: Low-Rank Adaptation (LoRA) via PEFT
• Dataset: ~500 high-quality instruction-response pairs
• Hardware: Trained on NVIDIA T4 GPU (Google Colab)

Data Pipeline

The dataset was curated using a hybrid approach:

1. Seed Problems: Real-world troubleshooting scenarios from StackOverflow and documentation.
2. Synthetic Expansion: Augmenting seed data using LLMs to cover edge cases.
3. Formatting: Structured in the Alpaca instruction-tuning format for better instruction following.

---

🚀 Usage

To use this adapter, you will need the transformers, peft, and torch libraries installed.

from transformers import AutoTokenizer, AutoModelForCausalLM
from peft import PeftModel
import torch

base_model_id = "mistralai/Mistral-7B-Instruct-v0.2"
lora_adapter_id = "Bharath5626/devops-lora-mistral" # Update with your specific HF path

tokenizer = AutoTokenizer.from_pretrained(base_model_id)
model = AutoModelForCausalLM.from_pretrained(
    base_model_id, 
    torch_dtype=torch.float16, 
    device_map="auto"
)
model = PeftModel.from_pretrained(model, lora_adapter_id)

# Inference Example
instruction = "Fix docker container exits immediately"
prompt = f"### Instruction:\n{instruction}\n\n### Response:\n"

inputs = tokenizer(prompt, return_tensors="pt").to("cuda")
outputs = model.generate(**inputs, max_new_tokens=150)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))