Nikhil1581/qwen3-4b-instruct-2507.Q4_K_M-excel-finetuning-1.2kdataset

Qwen3-4B-Instruct-2507 — Excel Fine-Tune (Q4_K_M)

A QLoRA fine-tuned version of Qwen/Qwen3-4B-Instruct-2507, specialized for Excel and spreadsheet tasks. Quantized to GGUF Q4_K_M for local deployment via Ollama or LM Studio.

---

Model Details

Property	Value
Base model	Qwen/Qwen3-4B-Instruct-2507
Parameters	4B
Fine-tune method	QLoRA (4-bit)
Dataset size	~1,200 samples
Quantization	GGUF Q4_K_M
File size	2.32 GB (34.5% reduction from 3.55 GB)
License	MIT

---

Evaluation Results

Manually evaluated on a 30-prompt benchmark spanning standard and advanced Excel tasks:

Category	Prompts	Accuracy
Standard (formulas, lookups, data analysis, basic VBA)	25	96%
Advanced (array formulas, VBA macros, financial modelling)	5	80%
Overall	30	~93%

Evaluation was performed via manual testing. A response was marked correct if it produced a working, usable formula or macro without requiring correction.

---

Training Performance

Step	Training Loss
25	1.0435
50	0.7119
75	0.6409
100	0.4136
125	0.4048
150	0.3663
175	0.2248
200	0.2225

• Loss reduction: 78.7% (1.04 → 0.22) over 200 steps
• Framework: Unsloth (2x faster training pipeline)
• Trainable parameters: ~132K (QLoRA adapters only)

---

Intended Use

This model is fine-tuned to assist with Excel and spreadsheet workflows:

• Writing and explaining Excel formulas (VLOOKUP, INDEX/MATCH, XLOOKUP, array formulas, etc.)
• Debugging broken formulas
• Data analysis with Excel (pivot tables, conditional formatting, data validation)
• VBA macro generation and explanation
• Converting between Excel functions and Python/Pandas equivalents
• Step-by-step spreadsheet task guidance

---

Quick Start

Ollama

ollama pull Nikhil1581/qwen3-4b-instruct-2507.Q4_K_M-excel-finetuning-1.2kdataset
ollama run Nikhil1581/qwen3-4b-instruct-2507.Q4_K_M-excel-finetuning-1.2kdataset

LM Studio

1. Search for Nikhil1581/qwen3-4b-instruct-2507 in the model browser
2. Download the Q4_K_M variant
3. Load and chat

llama.cpp

./llama-cli -m qwen3-4b-excel-q4_k_m.gguf \
  --chat-template chatml \
  -p "You are an Excel expert assistant." \
  -i

---

Example Prompts

Formula help:

User: How do I look up a value in column A and return the corresponding value from column C?

Debugging:

User: My VLOOKUP returns #N/A even though the value exists. Why?

VBA:

User: Write a VBA macro to loop through all sheets and highlight cells greater than 1000 in red.

Data analysis:

User: How do I calculate a running total in Excel without using a helper column?

---

Training Details

• Fine-tune type: QLoRA (4-bit quantized LoRA)
• LoRA rank: 16
• LoRA alpha: 32
• Target modules: q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj
• Dataset: ~1,200 Excel/spreadsheet instruction-response pairs
• Sequence length: 1,024 tokens
• Epochs: 3
• Steps: 200
• Optimizer: paged_adamw_8bit
• Hardware: T4 GPU (Google Colab)
• Framework: HuggingFace TRL + PEFT + Unsloth

---

Limitations

• Focused on Excel — general coding or math reasoning may be weaker than the base model
• Dataset is English-only
• Q4_K_M quantization may reduce precision on very complex multi-step formula chains
• Not tested on Google Sheets or LibreOffice Calc (though most formulas transfer)
• Evaluation was manual (25 standard + 5 advanced prompts) — not a formal benchmark

---

Recommended Inference Settings

temperature: 0.3
top_p: 0.9
repeat_penalty: 1.1
num_predict: 512

Low temperature (0.3) is recommended to keep formula syntax accurate.

---

Author

Nikhil1581 — HuggingFace Profile

---

Acknowledgements

• Qwen Team for the base model
• Unsloth for the 2x faster training framework
• HuggingFace TRL for the fine-tuning pipeline
• llama.cpp for GGUF quantization