metermind-ernie-gauge-lora

MeterMind: Industrial Gauge Reading with ERNIE-4.5-VL

Fine-tuned LoRA adapter for reading analog industrial gauges from images.

Model Description

MeterMind is a fine-tuned version of ERNIE-4.5-VL-28B-A3B that accurately reads analog gauge values from photographs. It was trained on synthetic industrial gauge images and achieves near-human accuracy.

• Base Model: unsloth/ERNIE-4.5-VL-28B-A3B-PT
• Fine-tuning Method: LoRA (Low-Rank Adaptation)
• Training Framework: Unsloth + TRL

Performance

Supported Gauge Types

• Standard pressure gauges (0-100 PSI)
• Glycerin-filled pressure gauges (0-100 PSI)
• Bimetal thermometers (0-220°F)

Training Details

Hyperparameters

• LoRA rank (r): 8
• LoRA alpha: 16
• Learning rate: 2e-4
• Training steps: 285 (1 epoch)
• Batch size: 2 (effective)
• Warmup steps: 10
• Scheduler: Linear
• Optimizer: AdamW 8-bit

Target Modules

q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj, fc1, fc2

Usage

from transformers import AutoProcessor, AutoModelForVision2Seq
from peft import PeftModel

# Load base model
base_model = AutoModelForVision2Seq.from_pretrained("unsloth/ERNIE-4.5-VL-28B-A3B-PT")
processor = AutoProcessor.from_pretrained("unsloth/ERNIE-4.5-VL-28B-A3B-PT")

# Load LoRA adapter
model = PeftModel.from_pretrained(base_model, "luliuzee/metermind-ernie-gauge-lora")

# Inference
prompt = "What is the reading on this gauge? Reply with only the numeric value."
# ... process image and generate

Intended Use

• Industrial inspection automation
• Digital rounds and operator logging
• Equipment monitoring systems
• Facility management applications

Limitations

• Trained on synthetic data; real-world performance may vary
• Best suited for circular analog gauges with clear markings
• Requires clear, well-lit images for optimal accuracy

Citation

@misc{metermind2024,
  title={MeterMind: Industrial Gauge Reading with Fine-Tuned ERNIE-4.5-VL},
  author={luliuzee},
  year={2024},
  publisher={HuggingFace}
}

Links

• Code Repository: GitHub (MeterMind)
• Hackathon: ERNIE AI Developer Challenge

Framework Versions

• PEFT: 0.18.0
• Transformers: 4.x
• Unsloth: Latest