Instructions to use iulio/FiscMind-Gemma4-26B-MoE with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- PEFT
How to use iulio/FiscMind-Gemma4-26B-MoE with PEFT:
from peft import PeftModel from transformers import AutoModelForCausalLM base_model = AutoModelForCausalLM.from_pretrained("google/gemma-4-26B-A4B-it") model = PeftModel.from_pretrained(base_model, "iulio/FiscMind-Gemma4-26B-MoE") - Notebooks
- Google Colab
- Kaggle
🏛️ FiscMind-Gemma4-26B-MoE (Frontier Romanian Fiscal & Accounting Assistant)
FiscMind-Gemma4-26B-MoE is an elite Romanian fiscal and accounting AI model adapted from Google's Gemma 4 (26B A4B MoE) architecture on high-performance NVIDIA A100-80GB SXM4 infrastructure via Modal.com.
It merges the parametric capacity of a 26-Billion parameter foundation model with the ultra-low latency and high token throughput of a 4-Billion active parameter sparse routing network, specialized for Romanian Statutory GAAP (OMFP 1802/2014), the Romanian Tax Code (Legea 227/2015), recent fiscal austerity packages (Legea 296/2023, OUG 115/2023), and mandatory digital compliance frameworks (RO e-Factura UBL 2.1, SAF-T D406).
⚡ Technical Highlights
| Attribute | Specification |
|---|---|
| Base Model | google/gemma-4-26B-A4B-it (Google DeepMind) |
| Architecture | Sparse Mixture-of-Experts (gemma4_unified) with 128 experts, top-8 routing |
| Active Parameters | ~4 Billion active parameters per token (25.82B total parameters) |
| Hardware Used | NVIDIA A100-SXM4-80GB (Modal.com Serverless Cluster) |
| Quantization at Training | 4-bit NF4 with native bfloat16 compute (BitsAndBytesConfig) |
| LoRA Config | $r=16$, $\alpha=32$, targeting language model linear projections (q, k, v, o, gate, up, down) |
| Training Steps | 100 steps (effective batch size 8) over 1,719 statutory curriculum items |
| Training Duration | 979.9 seconds (16.33 min) |
| Final Loss | 0.6011 (Mean token accuracy >99.9%) |
| Adapter Size | 35.51 MB SafeTensors |
📚 Domain Competencies & Curriculum
- OMFP 1802/2014 RO GAAP Accounting:
- Double-entry accounting monographies: purchases (371, 401, 4426), sales (4111, 707, 4427, 607), trade margins (378, 4428).
- Fixed assets & leasing (213, 167, 666, 6811, 2813).
- Foreign exchange variations (401, 5124, 7651, 6651).
- Closing operations & dividends (121, 129, 1061, 463, 457).
- Fiscal Code (Legea 227/2015 & Amending Legislation):
- Corporate Income Tax (Impozit pe Profit 16%, protocol 2%, sponsorships credit).
- New Minimum Turnover Tax (IMCA 1% for entities exceeding 50,000,000 EUR turnover).
- Fiscal loss carryforward limit (70% cap over 5 consecutive fiscal years).
- VAT reverse charge (Taxare inversă 4426 = 4427) & Intra-community acquisitions (AIC/LIC).
- Payroll & Extrasalary Benefits (OUG 115/2023 & Legea 296/2023):
- IT sector tax exemption cap (10,000 lei gross salary ceiling).
- Meal vouchers mandatory CASS (10% health contribution retention).
- 33% monthly tax-free extrasalary benefits threshold.
- Digital Tax Compliance:
- RO e-Factura (5 calendar days transmission deadline, 15% sanction for unregistered invoices).
- RO e-Transport (UIT code validity of 5 calendar days, 500 kg / 10,000 lei thresholds).
- SAF-T (D406) (General ledger taxonomy mapping, monthly/quarterly obligations).
🚀 Quickstart: Inference with Hugging Face & PEFT
import torch
from transformers import AutoTokenizer, AutoModelForCausalLM
from peft import PeftModel
BASE_MODEL = "google/gemma-4-26B-A4B-it"
ADAPTER_REPO = "iulio/FiscMind-Gemma4-26B-MoE"
tokenizer = AutoTokenizer.from_pretrained(BASE_MODEL)
base_model = AutoModelForCausalLM.from_pretrained(
BASE_MODEL,
device_map="auto",
dtype=torch.bfloat16,
)
model = PeftModel.from_pretrained(base_model, ADAPTER_REPO)
model.eval()
sys_msg = (
"Esti FiscMind AI, expert contabil CECCAR si consultant fiscal CCF conform "
"legislatiei din Romania actualizate 2025/2026. Raspunde precis cu temei legal si formule contabile complete."
)
prompt = f"""<start_of_turn>user
{sys_msg}
Care este monografia contabila pentru inregistrarea dividendelor interimare trimestriale si ce impozit se retine?<end_of_turn>
<start_of_turn>model
"""
inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
with torch.no_grad():
outputs = model.generate(
**inputs,
max_new_tokens=512,
temperature=0.2,
do_sample=True,
)
print(tokenizer.decode(outputs[0][inputs.input_ids.shape[1]:], skip_special_tokens=True))
🛡️ Safety, Guardrails & Human-in-the-Loop Standard
- Strict Non-Autonomous Final Approval: All accounting and fiscal recommendations require human chartered accountant (CECCAR/CCF) verification before statutory declaration submission or general ledger booking.
- Audit Trail: Every monetary suggestion maintains strict double-entry balance: $$\sum \text{Debit} \equiv \sum \text{Credit}$$
- Data Privacy: No client confidential data or real company tax numbers are committed to public models.
📜 Citation & Credits
Developed by Iulian Amaricai as part of the FiscMind AI Romanian Statutory Accounting Assistant initiative. Base model courtesy of Google DeepMind (Gemma 4 Family).
- Downloads last month
- 12