---
language:
- en
- zh
license: apache-2.0
library_name: transformers
tags:
- industrial
- maintenance
- fault-diagnosis
- sop-generation
- qwen
base_model: Qwen/Qwen3-4B
pipeline_tag: text-generation
---

# emsLLM: Industrial Equipment Fault Diagnosis & Maintenance Assistant

## Model Summary

**emsLLM** is a Large Language Model (LLM) fine-tuned specifically for the Electronic Manufacturing Services (EMS) and industrial maintenance sectors. Built upon the **Qwen3-4B** architecture, this model has been trained on equipment operation manuals, troubleshooting guides, and standardized maintenance documents. It is designed to assist engineers with **equipment fault diagnosis** and **SOP (Standard Operating Procedure) generation**.

This is a **Merged Version**, meaning the fine-tuned LoRA weights have been merged into the base model. It can be loaded directly for inference without requiring additional adapters, ensuring easy deployment and integration.

## Evaluation & Performance

We evaluated **emsLLM** against state-of-the-art general-purpose models (**Llama-3.3-70B** and **Qwen3-32B**) using an "LLM-as-a-Judge" approach. The evaluation focused on industrial fault diagnosis accuracy and inference speed.

### 1. Diagnosis Accuracy
Despite their significantly smaller parameter count, our fine-tuned models (trained on public technical documents) outperformed larger base models in retrieving and generating correct maintenance solutions.

| Model | Public Test Set Accuracy | Private Test Set Accuracy |
| :--- | :---: | :---: |
| **emsLLM-8B (Ours)** | **95.0%** | **100.0%** |
| **emsLLM-4B (Ours)** | **95.0%** | **96.0%** |
| Qwen3-32B (Base) | 90.0% | 92.0% |
| Llama-3.3-70B (Base) | 87.5% | 92.0% |

> **Key Finding:** The **emsLLM-8B** achieves **95% accuracy** on public technical queries and **100%** on private domain tasks, surpassing the Llama-3.3-70B base model while using significantly fewer resources.

### 2. Inference Speed
For industrial applications requiring real-time response, our models offer ultra-low latency.

| Model | Time to First Token (TTFT) | Time Per Output Token (TPOT) |
| :--- | :---: | :---: |
| **emsLLM-4B** | **30.46 ms** | **3.20 ms** |
| **emsLLM-8B** | **58.92 ms** | **6.40 ms** |
| Qwen3-32B | 253.49 ms | 20.91 ms |
| Llama-3.3-70B | 428.14 ms | 36.02 ms |

> **Key Finding:** The **emsLLM-8B** provides an approximate **6x speedup** in generation speed (TPOT) compared to Llama-3.3-70B, making it highly suitable for edge deployment and rapid interaction.

## Key Features

*   **Fault Diagnosis Assistant**: Provides analysis of potential causes and maintenance suggestions for common production line equipment (e.g., robotic arms, dispensing machines).
*   **SOP Content Generation**: Understands industrial documentation logic and assists in drafting standardized SOPs, including operational procedures and responsibility assignments.
*   **Domain Optimization**: Enhanced understanding of maintenance terminology and logic compared to generic models, offering responses more aligned with engineering field requirements.
*   **Ready-to-Deploy**: Weights are fully merged, supporting direct loading with standard inference frameworks (e.g., Hugging Face Transformers, vLLM).

## Quick Start

Since the weights are already merged, you can load the model just like any standard Transformers model:

```python
from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

# Replace with your actual Hugging Face model ID
model_id = "Your-Username/emsLLM" 

# 1. Load Tokenizer
tokenizer = AutoTokenizer.from_pretrained(model_id)

# 2. Load Model
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    torch_dtype=torch.bfloat16,
    device_map="auto"
)

# 3. Example: Fault Diagnosis
messages = [
    {"role": "system", "content": "You are a professional industrial equipment maintenance assistant."},
    {"role": "user", "content": "The dispensing machine on the production line has unstable output. What are the possible causes and solutions?"}
]

text = tokenizer.apply_chat_template(
    messages,
    tokenize=False,
    add_generation_prompt=True
)

model_inputs = tokenizer([text], return_tensors="pt").to(model.device)

generated_ids = model.generate(
    model_inputs.input_ids,
    max_new_tokens=512,
    temperature=0.1  # Low temperature is recommended for stability in industrial tasks
)

response = tokenizer.batch_decode(generated_ids, skip_special_tokens=True)[0]
print(response)

```

## **Contact info**
- Email: fred_tung@pegatroncorp.com