Nano Language Models

Building 6SigmaMind: How I Tuned a Tiny Model to Understand Excel

Most AI projects start with massive models.
6SigmaMind started with a challenge:

“How smart can a tiny model become if you teach it just one skill extremely well?”

That skill:
Turning everyday English into real, working Excel formulas.

The result is a lightweight assistant powered by a 1.7B-parameter model — small enough to run inside a Hugging Face Space, yet surprisingly capable at one job.

Today, I want to share how I built it, what worked, and where it’s heading next.

👉 Try 6SigmaMind live: https://huggingface.co/spaces/benkemp/6SigmaMindv2

🌱 The Starting Point: A Small Model With Big Ambitions

I began with HuggingFaceTB/SmolLM2-1.7B —
a compact, efficient SLM that loads fast and responds instantly even on CPU.

On paper, 1.7B parameters is tiny.
But I chose it deliberately because:

  • small models run anywhere
  • they’re fun to experiment with
  • they respond fast
  • they teach you what’s possible without giant compute
  • and they’re perfect for task-specific specialization

I didn’t want an “everything AI.”
I wanted a precision tool — something like a pocket calculator for formulas.

🧩 Step 1 — Designing the Prompt Brain

The brain of 6SigmaMind isn’t just the model — it’s the instruction.
Tiny models need clear direction.

The system prompt I designed tells the model:

  • always focus on Excel logic
  • prefer returning a single formula
  • avoid long explanations
  • behave consistently across all prompt styles
  • format responses cleanly

Then I tested hundreds of variations like:

  • “Write an Excel formula that…”
  • “Give me the formula for…”
  • “If A2 > 100…”
  • “Calculate standard deviation…”
  • “Return the last non-empty value…”

This gave the model its “voice”:
fast, minimal, formula-first.

🔍 Step 2 — Building a Test Suite of Real Excel Questions

To shape the model’s behavior, I created a large collection of real Excel tasks:

  • SUMIFS
  • COUNTIFS
  • XLOOKUP
  • INDEX/MATCH
  • text extraction
  • ranges
  • date math
  • statistics
  • filtering
  • ranking
  • regression basics

These created a natural benchmark.

If a prompt like
“Sum C where B is ‘Closed’”
didn’t immediately produce:

=SUMIFS(C:C, B:B, "Closed")

…then I knew something needed tuning.

Small models get better through repetition.
Testing is training.

🛠️ Step 3 — Making It Work in a Public Demo

The Hugging Face Space needed to be:

  • fast
  • clean
  • simple
  • fun

So I built:

  • a Gradio UI
  • sliders for temperature and token length
  • examples visitors can click
  • strict formula formatting
  • text-only inference (so it runs anywhere)

The result:
You type a sentence → the model replies instantly.

No waiting.
No login.
No friction.

A tool people actually want to play with.

⚡ What Worked Surprisingly Well

✔ Logical formulas

IF, AND, OR — tiny models handle these beautifully.

✔ Simple lookups

XLOOKUP and INDEX/MATCH are surprisingly strong.

✔ Summaries

Counting, summing, averaging — very reliable.

✔ Text extraction

LEFT, MID, RIGHT, SEARCH — consistently good.

🧪 Where Tiny Models Struggle (and That’s Okay)

Tiny models still get challenged by:

  • tricky statistical formulas
  • multi-condition nested logic
  • complex filters
  • rare Excel functions
  • unusual phrasing or typos
  • long multi-step tasks

And that’s exactly why the experiment is interesting.

If everything worked perfectly, I’d be building another GPT-4 clone.
Instead, we’re pushing the boundary of “how smart can small models become?”

🧠 Why Build 6SigmaMind in the First Place?

Because the future of AI isn’t just about giant models.
It’s about:

  • speed
  • focus
  • accessibility
  • specialization
  • local-friendly models
  • tools that do one job extremely well

6SigmaMind is a small proof-of-concept of that philosophy:
A tiny model + a precise domain = surprising power.

🚀 What’s Next for 6SigmaMind?

Here’s what’s coming soon:

🔹 Formula-only mode

For benchmark testing and evaluation.

🔹 Training on Excel for Statistics

STDEV, VAR, CORREL, T.TEST, regression, confidence intervals.

🔹 Dataset expansion

1000+ instruction–formula pairs.

🔹 Fine-tuning

Turning the model into a true Excel specialist.

🔹 Model comparison dashboard

Phi-3 Mini vs Qwen-3B vs SmolLM2 vs Gemma-2B.

🔹 Embeddable widget

Use 6SigmaMind directly in websites, apps, or notebooks.

This is just the beginning.

🎯 Try It Yourself — The Fun Starts Here

👉 Test 6SigmaMind live: https://huggingface.co/spaces/benkemp/6SigmaMindv2
Try your real Excel problems.
Try weird ones.
Try breaking it.

Every prompt teaches us something new about what small models can do.

And together, we’ll make the next version even smarter.

Latest Articles