covaga/eplan-2027-lora-vision

EPLAN Electrical Symbol Recognition — Qwen3.5-9B Vision LoRA

Vision LoRA adapter on Qwen/Qwen3.5-9B that looks at an EPLAN / IEC / NFPA electrical schematic symbol and describes it: shape, structure, type of contact/function, and EPLAN symbol naming. Trained on 24k real EPLAN symbol images — part of the open EPLAN AI toolchain by Covaga · github.com/covagashi.

🚀 Quickstart

# pip install "transformers>=5" unsloth peft pillow
from unsloth import FastVisionModel
from peft import PeftModel
from PIL import Image

model, processor = FastVisionModel.from_pretrained(
    "Qwen/Qwen3.5-9B", load_in_4bit=False, load_in_16bit=True,
    max_seq_length=2048)
model = PeftModel.from_pretrained(model, "covaga/eplan-2027-lora-vision")
FastVisionModel.for_inference(model)

img = Image.open("symbol.png").convert("RGB")
msgs = [{"role": "user", "content": [
    {"type": "image", "image": img},
    {"type": "text",
     "text": "Identify this EPLAN electrical symbol and describe its shape."}]}]
inputs = processor.apply_chat_template(
    msgs, tokenize=True, return_dict=True,
    return_tensors="pt", add_generation_prompt=True).to(model.device)
out = model.generate(**inputs, max_new_tokens=300)
print(processor.decode(out[0][inputs["input_ids"].shape[1]:],
                       skip_special_tokens=True))

Requires transformers>=5. Needs ≥24 GB VRAM for BF16 (or quantize / use the merged GGUF ecosystem below).

🧠 What it does

• Sees an EPLAN symbol image (blue schematic glyphs, ~400×400)
• Describes its visual form accurately: lines, contacts, circles, diagonals, connection points — verified qualitatively on held-out symbols
• Identifies symbol type (fuse, pushbutton, terminal, transformer…) and attempts the EPLAN short_name + variant

Verified results (24 unseen test symbols)

Aspect	Result
Shape/structure description	✅ Accurate — correct primitives, layout, connection points
Symbol short_name recall	⚠️ 1/24 — arbitrary ID memorization is hard; use visual RAG for naming (below)

Example of what it produces — given a mushroom pushbutton symbol:

"A vertical line segment. A right-pointing triangle attached to the right side… a horizontal dashed line extending from the tip of the triangle."

📚 Dataset

covaga/electrical-symbols-dataset — 33,502 EPLAN symbol images with short_name, catalog number, IEC/NFPA description, variant and visual_description. Trained on 24k samples, split by short_name so test symbols were never seen.

🔌 Recommended: pair with visual RAG for exact naming

Exact short_name recall needs retrieval, not memorization: embed the query image (CLIP/SigLIP), look up nearest neighbors in the symbol dataset, feed the real names as candidates. Related tooling:

• covagashi/eplan-rag-mcp — EPLAN MCP server, verified 2027 action registries
• covagashi/eplan-cloudflare-rags — hosted RAG over official EPLAN 2027 docs
• covagashi/eplan-development-skill — curated EPLAN dev reference for agents

🧩 Related models

• covaga/eplan-2027-lora-v3 — text adapter: EPLAN 2027 API Q&A + C# scripting (EN/ES)
• covaga/eplan-2027-9b-merged — merged BF16 text model
• covaga/eplan-2027-9b-merged-GGUF — Q4_K_M for local use

🏋️ Training details

Setting	Value
Method	Vision LoRA (Unsloth FastVisionModel), BF16
Layers tuned	vision encoder + language attention/MLP (102M params, 1.07%)
Samples	24,000 image QA pairs, 1 epoch (1,500 steps, eff. batch 16)
Final loss	0.50 avg · 0.24 last quarter
Hardware	A100 80GB, 2.3 h, 19.5 GiB VRAM peak

⚠️ Limitations

• Weak on exact symbol IDs (1/24) — pair with image retrieval for production naming.
• Long answers can burn the token budget on shape analysis before naming; give it ≥300 tokens or ask for the name first.
• Isolated symbols only — it was not trained on full schematic pages.
• Dataset noise exists (some auto-generated descriptions are garbled).
• Not a replacement for EPLAN: verify against the official symbol library.
• Base model Qwen/Qwen3.5-9B license terms apply.

📝 Citation

@misc{eplan-2027-lora-vision,
  author  = {Lopez, Christian D.},
  title   = {EPLAN Electrical Symbol Recognition — Qwen3.5-9B Vision LoRA},
  year    = {2026},
  url     = {https://huggingface.co/covaga/eplan-2027-lora-vision}
}

---

Part of the covaga.xyz open EPLAN tooling — free Spanish EPLAN scripting guide, RAG endpoints, datasets and models.