EPLAN Electric P8 2027 API Assistant — merged model (Qwen3.5-9B, BF16)
Full-precision BF16 merged model (18.8 GB): Qwen3.5-9B + the
eplan-2027-lora-v3
adapter baked in — load it directly, no PEFT needed. Trained for EPLAN
Electric P8 2027 API questions, C# script generation for the EPLAN
scripting engine, error diagnosis, and refusal of nonexistent APIs.
Bilingual English + Spanish.
🚀 Quickstart
# pip install "transformers>=5" torch
from transformers import AutoModelForCausalLM, AutoTokenizer
tok = AutoTokenizer.from_pretrained("covaga/eplan-2027-9b-merged")
model = AutoModelForCausalLM.from_pretrained(
"covaga/eplan-2027-9b-merged", torch_dtype="bfloat16", device_map="auto")
prompt = "Write a C# script for EPLAN 2027 that exports every schematic page to PDF."
ids = tok(tok.apply_chat_template(
[{"role": "user", "content": prompt}],
tokenize=False, add_generation_prompt=True), return_tensors="pt").to(model.device)
print(tok.decode(model.generate(**ids, max_new_tokens=512)[0][ids["input_ids"].shape[1]:]))
Requires
transformers>=5— Qwen3.5 is not supported on v4.x.
💻 Requirements
| Setup | Minimum | Comfortable |
|---|---|---|
| GPU | ≥24 GB VRAM | A100 / RTX 3090–4090 |
| CPU | ~24 GB+ RAM | 32 GB+ (very slow) |
For laptops / ≤16 GB machines use the GGUF Q4_K_M build (5.6 GB):
covaga/eplan-2027-9b-merged-GGUF.
This merged model is also the right starting point for your own quantization
(GGUF, AWQ, GPTQ) or further fine-tuning — prefer it over re-merging adapters.
🧠 What it does
- Answers questions about the EPLAN Electric P8 2027 API (
Eplan.EplApi.*) - Generates C# scripts for the EPLAN scripting engine
- Explains compile errors (CS0234…) and runtime issues (message-loop hangs,
Netuassembly pitfalls) - Refuses (most of the time) APIs that don't exist instead of fabricating
- Works in English and Spanish
📊 Benchmarks (curated 44-question eval)
| Arm | Valid identifiers | Invented IDs | Correct refusals |
|---|---|---|---|
| Base Qwen3.5-9B | 43% | 230 | 1/9 |
| This model (v3 LoRA merged) | 63% | 47 | 4/9 |
Clean held-out test split (60 questions, never seen in training): 63% valid
identifiers, 60/60 expected IDs, 0 constraint violations. Full comparison
including the more conservative v4 adapter:
eplan-2027-lora-v3 ·
eplan-2027-lora-v4.
📚 Dataset
Trained on covaga/eplan-2027-api-qa:
~19k Q&A pairs generated from the official EPLAN 2027 API documentation,
split by source document to prevent leakage, plus 246 curated examples of
refusal/ambiguity/error-diagnosis verified against the live docs.
Evaluation data: covaga/eplan-2027-eval.
🔌 Make it better — ground it with real docs
Pairing with retrieval is the single biggest quality win (in our eval, retrieval alone lifted refusal accuracy on invented-API traps from 1/9 to 4/9):
covagashi/eplan-cloudflare-rags— hosted RAG over the official EPLAN 2027 docs (rag2027.covaga.xyz)covagashi/eplan-rag-mcp— MCP server with verified 2027 action registriescovagashi/eplan-development-skill— curated EPLAN development reference for agents
⚠️ Limitations
- Still fabricates plausible-sounding API names in ~half of trap cases — always verify against the official docs.
- Text-only: the vision tower of Qwen3.5-9B was not trained — no EPLAN screenshot understanding.
- Identifier exists ≠ code correct — compile generated scripts in a real EPLAN installation before production use.
- Synthetic dataset may carry upstream errors; base model license terms apply.
📝 Citation
@misc{eplan-2027-9b-merged,
author = {Lopez, Christian D.},
title = {EPLAN 2027 API Assistant — Qwen3.5-9B merged},
year = {2026},
url = {https://huggingface.co/covaga/eplan-2027-9b-merged}
}