Small Language Models (SLMs) aren’t just capable of generating text — they can translate natural language instructions into complex Excel formulas with surprising accuracy. With the right prompting techniques and a well-designed training dataset, an SLM can understand requests like:

• “Find all rows where Sales > 10,000 and Region = ‘West’.”
• “Sum the last 7 entries in column B only if column A equals ‘Active’.”
• “Extract the domain name from an email address.”

…and convert these commands into fully working Excel formulas.

This article shows how SLMs interpret human requests, how prompts should be structured, and how to build an Excel Formula Assistant model using your TinySLM, SmolLM, Granite Nano, or any SLM you’re fine-tuning.

1. Why Excel Formulas Are the Perfect Use Case for Small Models

Excel formulas are:

• Structured
• Predictable
• Deterministic
• Short and syntactically narrow

This makes them ideal for SLMs (350M–1.7B), which excel at tasks requiring:

• Logical pattern recognition
• Formula transformation
• Symbolic reasoning
• Limited context windows

You don’t need GPT-4 for SUMIFS, VLOOKUP, or FILTER.
A specialized SLM can outperform larger models for this niche.

2. Core Prompt Structure for Excel Formula Generation

For consistent results, SLMs respond best to a simple, strict prompt format.

Recommended Prompt Template

Convert the following instruction into a valid Excel formula.
Instruction: "<user_instruction>"
Only return the Excel formula.

Example:

Convert the following instruction into a valid Excel formula.
Instruction: "Sum all values in column E where column B equals 'North'."
Only return the Excel formula.

Expected SLM output:

=SUMIF(B:B,"North",E:E)

This structure ensures:

• Clear intent
• No extra text
• No explanation
• Predictable output

3. Example Inputs & Outputs for Training

These examples help shape your fine-tuning dataset.

Input

Instruction: "Filter rows where F > 100 and G = 'Yes'."

Output

=FILTER(A:Z,(F:F>100)*(G:G="Yes"))

Input

Instruction: "Find the average of C2:C100 but only for rows where column D equals 'Closed'."

Output

=AVERAGEIF(D2:D100,"Closed",C2:C100)

Input

Instruction: "Extract the last 4 characters from each value in column B."

Output

=RIGHT(B:B,4)

These examples help give the model confidence.
Your dataset should contain hundreds to thousands of such pairs.

4. How SLMs Parse Natural Language into Formula Logic

SLMs break down your input into three steps:

Step 1 — Identify the action

Sum, filter, count, average, lookup, extract…

Step 2 — Identify the conditions

“>100”, “equals Closed”, “not empty”, “contains @gmail”

Step 3 — Map the logic to a formula pattern

• SUMIF / SUMIFS
• FILTER
• INDEX + MATCH
• VLOOKUP / XLOOKUP
• LEFT / RIGHT / MID

This mapping is learnable through pattern repetition —
making SLMs incredibly efficient at formula generation.

5. Improving Accuracy with Prompt Engineering

To reduce hallucinations:

a) Force output constraints

Only return one Excel formula with no explanation.

b) Add examples

Few-shot prompting:

Instruction: "Get the first word from each cell in A:A."
Output: =LEFT(A:A, FIND(" ",A:A)-1)

Instruction: "Sum all values in C where B equals 'X'."
Output: =SUMIF(B:B,"X",C:C)

Instruction: "<new_instruction>"
Output:

c) Add formatting rules

Use uppercase function names.
Do not explain.
Return nothing but the formula.

SLMs follow these rules surprisingly well.

6. Fine-Tuning Your Own Excel Formula SLM

To fine-tune a model like SmolLM-360M or Granite 4.0 Nano:

Dataset Format (JSONL)

{"instruction": "Sum all values in column C.", "output": "=SUM(C:C)"}
{"instruction": "Filter rows where G = 1.", "output": "=FILTER(A:Z, G:G=1)"}

Ideal size:

• 5,000 entries = usable model
• 20,000 entries = very accurate
• 50,000+ entries = near-perfect

Train using:

• QLoRA
• PEFT
• Hugging Face TRL
• 4-bit quantization

Within 1–3 hours, you get a powerful Excel assistant running fully local.

7. Deploying the Excel Formula SLM

You can deploy the model into:

• Gradio (Hugging Face Space)
• FastAPI server
• VS Code extension
• Excel add-in
• Local Python desktop app

A simple Gradio interface lets users type:

“Get unique values from column A”

…and instantly receive:

=UNIQUE(A:A)