A fine-tuned version of Qwen/Qwen2.5-3B-Instruct trained on the IRS tax code using a 3-stage training pipeline. This repository contains the quantized GGUF model for local inference via llama.cpp, Ollama, or compatible runtimes.
Model Details
Property
Value
Base Model
Qwen/Qwen2.5-3B-Instruct
Parameters
~3 Billion
Quantization
q8_0 (8-bit, high quality)
Format
GGUF
License
Apache 2.0
Training Domain
IRS Tax Code / US Federal Tax Law
Training Pipeline
This model was produced by a 3-stage fine-tuning pipeline applied to Qwen 2.5 3B Instruct:
Stage 1 — Supervised Fine-Tuning (SFT)
The base instruct model was fine-tuned with LoRA adapters on a curated dataset of IRS tax code passages, definitions, and question-answer pairs. This stage establishes domain knowledge and instills the correct response format for tax-related queries.
Stage 2 — Direct Preference Optimization (DPO)
Using pairs of preferred and rejected responses to tax questions, DPO training aligned the model to produce accurate, well-structured answers that reflect correct tax law interpretation over hallucinated or misleading responses.
Stage 3 — Group Relative Policy Optimization (GRPO)
GRPO reinforcement learning was applied to further optimize response quality using a reward model that evaluated accuracy, completeness, and citation of relevant tax code sections. This stage improves the model's ability to reason through multi-step tax scenarios.
Files
File
Description
qwen25-tax-3b-q8_0.gguf
q8_0 quantized GGUF for local inference
LoRA Adapters
The raw LoRA adapter weights for all three training stages are available in the companion repository: dennisonb/qwen25-tax-3b-adapters
Usage
With Ollama
bash
1ollama create qwen25-tax-3b -f Modelfile
2ollama run qwen25-tax-3b "What is the standard deduction for a single filer in 2024?"
With llama.cpp
./llama-cli -m qwen25-tax-3b-q8_0.gguf -p "What are the income limits for contributing to a Roth IRA?" -n 512
Intended Use
This model is intended as a research and educational tool for exploring IRS tax code information. It is not a substitute for professional tax advice. Always consult a qualified tax professional for your specific situation.
Limitations
Trained primarily on IRS publications and the US Internal Revenue Code
Knowledge cutoff tied to the training data; tax law changes frequently
May not reflect the most recent legislative updates
Should not be used for actual tax filing decisions without professional review
