You can deploy this model and start earning money today!

arghyachaks/
GTOP-RL-Ministral-3B

Name: GTOP-RL-Ministral-3B
Availability: InStock

safetensors

reinforcement-learning

unsloth/Ministral-3-3B-Instruct-2512

finetune

cc-by-4.0

Views

No views yet

🌍 GTOP-RL: Global Tax Optimization Platform

Aligning Small Language Models with International Tax Law using Group Relative Policy Optimization (GRPO)

📖 Abstract

Large Language Models often struggle with reasoning tasks that require strict adherence to complex regulatory frameworks. GTOP-RL is a specialized 3B-parameter agent tuned for compliance with the OECD BEPS Pillar Two framework.

By modeling tax planning as a constrained resource allocation game and using Group Relative Policy Optimization (GRPO), we transformed a generic text generator into a compliance-aware specialist.

📊 Key Results

Compared to the baseline Ministral-3B model, GTOP-RL demonstrates a 7.1x improvement in tax savings while maintaining 98% regulatory compliance.

Model	Win Rate	Avg Savings (€M)	Compliance Error
Baseline (Untrained)	8.0%	€2.10M	24.0%
GTOP-RL (Ours)	34.0%	€15.10M	2.0%

🚀 Quick Start

You can run this model using the unsloth library for 2x faster inference.

python

from unsloth import FastVisionModel

# 1. Load Model
model, tokenizer = FastVisionModel.from_pretrained(
    "arghyachaks/GTOP-RL-Ministral-3B",
    load_in_4bit=True,
)
FastVisionModel.for_inference(model)

# 2. Run Inference
messages = [
    {"role": "user", "content": "Analyze the tax liability for a subsidiary in Ireland with €50M revenue..."}
]
inputs = tokenizer.apply_chat_template(messages, return_tensors="pt").to("cuda")
outputs = model.generate(inputs, max_new_tokens=500)
print(tokenizer.decode(outputs[0]))