ERNIE 4.5 FWD Triage

Inspiration

Last month, my dad noticed a rattle in the front of his FWD sedan. It only happened over bumps, and he wasn’t sure if it was minor or serious. Most people don’t know, and that uncertainty leads to stress, hesitation, or bad decisions.

I tested the base ERNIE 4.5 and found the same flaw every time: long, vague checklists with no clear direction. It knew symptoms, but didn’t think like a real mechanic.

That made me realize: good diagnostics aren’t about listing everything. They’re about pattern recognition, confidence, and useful next steps. So I built a model that acts like a mechanic in your phone. You describe the noise, and it gives you the most likely failure. Not a full inspection, just enough clarity to act.

What It Does

It’s a fine-tuned ERNIE 4.5 (21B) model, trained with LoRA, that takes natural language descriptions of front-end noise symptoms in FWD sedans and returns top suspected causes: CV joints, wheel bearings, struts, loose shields, and so on.

It’s trained to commit to a diagnosis, not hedge like general LLMs. If a user describes a symptom outside of scope, it says so clearly.

How I Built It

• Model: ERNIE 4.5, fine-tuned using QLoRA (4-bit) on 1000+ structured symptom/diagnosis pairs
• Framework: Used Unsloth for efficient training and ChatML formatting
• Dataset: Manually constructed and bucketed samples covering common FWD sedan front-end issues
• Infrastructure: Trained in Colab with GPU using lightweight adapters and full reproducibility

Challenges I Ran Into

• Sourcing high-quality, diverse symptom phrasing without repetition
• Preventing the model from hallucinating or over-hedging
• Making the model say “I don’t know” confidently when outside domain

Accomplishments I'm Proud Of

• Model confidently narrows down probable causes from plain English descriptions
• Clearly responds when out-of-domain
• Fully reproducible setup using only open-source tools and Hugging Face
• Helped my dad identify a loose heat shield and save a shop trip

What I Learned

• Fine-tuning even a base model like ERNIE 4.5 becomes powerful with tight domain focus
• LoRA makes prototyping on billion-scale models practical even without heavy infra
• The power of narrow agents over generalist chatbots in vertical use-cases

What's Next for ERNIE 4.5 FWD Triage

• Expanding dataset to cover:
  • Engine knocking
  • Brake grinding
  • Transmission lag
• Longer-term: A full multi-modal app with audio inputs

Limitations

This is not a replacement for a shop inspection. It's a triage tool. Think of it like texting a mechanic friend — they give you a lead, but you still go to the shop if something’s really broken.

It’s trained strictly on front-end noise patterns in FWD sedans. For other domains, it will politely refuse or admit uncertainty.

Thanks for checking this out. If it helps even a few people avoid confusion at the mechanic’s — worth it.

Built With

• ernie
• huggingface
• python
• pytorch
• transformers
• unsloth