Nemotron 3 Nano 4B — Power Apps Formula Expert

A fine-tuned version of NVIDIA Nemotron 3 Nano 4B specialized in Microsoft Power Apps formula generation, error diagnosis, and SharePoint integration patterns.

Model Details

Attribute	Details
Developed by	DevNexAI
Base model	NVIDIA Nemotron 3 Nano 4B (Hybrid Mamba-Transformer MoE)
Fine-tuning method	QLoRA (4-bit) via Unsloth
LoRA rank	16
LoRA alpha	32
Training hardware	NVIDIA A100-SXM4-40GB (Google Colab Pro)
Precision	BF16
License	Apache 2.0
Languages	English, Spanish

What This Model Does

This model is trained to assist Power Apps developers with:

Formula generation — Describe what you need in natural language, get the correct Power Apps formula
Error diagnosis — Paste your broken formula, get the root cause and fix
SharePoint patterns — Delegation-aware queries, Patch operations, Choice/Lookup column handling
Regional syntax — Understands separator differences between English (,) and Spanish (;) environments
Best practices — Named Formulas, Concurrent loading, batch Patch, role-based visibility

Training Data

The model was fine-tuned on 202 curated examples across 5 categories:

Category	Examples	%
Error/Debugging	71	35.1%
Formula generation	61	30.2%
Patterns/How-to	56	27.7%
SharePoint specific	10	5.0%
Explanation	4	2.0%

Dataset: Devnexai/Nemotron_PowerApps_FineTunning

Usage

With Unsloth (recommended)

from unsloth import FastLanguageModel

model, tokenizer = FastLanguageModel.from_pretrained(
    "Devnexai/nemotron3-powerapps-formulas",
    load_in_4bit=True,
    trust_remote_code=True,
)
FastLanguageModel.for_inference(model)

messages = [
    {"role": "system", "content": "Eres un experto en Power Apps. Genera fórmulas correctas, explica errores y ayuda a resolver problemas de sintaxis."},
    {"role": "user", "content": "Filtra la lista Empleados donde el departamento sea Ventas y el estatus sea Activo"},
]

inputs = tokenizer(
    tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True),
    return_tensors="pt"
).to("cuda")

outputs = model.generate(**inputs, max_new_tokens=512, temperature=0.7, top_p=0.9)
print(tokenizer.decode(outputs[0][inputs["input_ids"].shape[-1]:], skip_special_tokens=True))

With vLLM (API server)

vllm serve unsloth/NVIDIA-Nemotron-3-Nano-4B-BF16 \
    --enable-lora \
    --lora-modules powerapps=Devnexai/nemotron3-powerapps-formulas

Example Prompts

Prompt	Category
`Filtra pedidos con monto mayor a 10000 y estatus Pendiente`	Formula generation
`Error: Filter(Lista, Status = "Activo") me marca error de tipo`	Error diagnosis
`Cómo hago una búsqueda delegable en SharePoint por múltiples campos`	SharePoint pattern
`Mi Patch no guarda datos, no da error pero no se actualiza`	Silent error debugging
`Copié una fórmula de internet y no funciona: If(x > 1, "Sí", "No")`	Regional syntax

Limitations

Trained on 202 examples — increasing to 500+ will improve quality
Optimized for SharePoint as datasource; Dataverse patterns are limited
GGUF export not available yet (Mamba-Transformer hybrid not supported by llama.cpp)
Requires Linux or WSL2 for local deployment (mamba_ssm dependency)

About DevNexAI

DevNexAI is an AI consulting and products venture focused on the Mexican and LATAM enterprise market, specializing in applied AI for enterprise platforms.