Instructions to use iulio/FiscMind-Qwen38-27B-CoT with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use iulio/FiscMind-Qwen38-27B-CoT with PEFT:

from peft import PeftModel
from transformers import AutoModelForCausalLM

base_model = AutoModelForCausalLM.from_pretrained("Qwen/Qwen3.8-27B")
model = PeftModel.from_pretrained(base_model, "iulio/FiscMind-Qwen38-27B-CoT")

Transformers

How to use iulio/FiscMind-Qwen38-27B-CoT with Transformers:

# Use a pipeline as a high-level helper
from transformers import pipeline

pipe = pipeline("text-generation", model="iulio/FiscMind-Qwen38-27B-CoT")
messages = [
    {"role": "user", "content": "Who are you?"},
]
pipe(messages)

# Load model directly
from transformers import AutoModel
model = AutoModel.from_pretrained("iulio/FiscMind-Qwen38-27B-CoT", device_map="auto")

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use iulio/FiscMind-Qwen38-27B-CoT with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "iulio/FiscMind-Qwen38-27B-CoT"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "iulio/FiscMind-Qwen38-27B-CoT",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/iulio/FiscMind-Qwen38-27B-CoT

SGLang

How to use iulio/FiscMind-Qwen38-27B-CoT with SGLang:

Install from pip and serve model

# Install SGLang from pip:
pip install sglang
# Start the SGLang server:
python3 -m sglang.launch_server \
    --model-path "iulio/FiscMind-Qwen38-27B-CoT" \
    --host 0.0.0.0 \
    --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "iulio/FiscMind-Qwen38-27B-CoT",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker images

docker run --gpus all \
    --shm-size 32g \
    -p 30000:30000 \
    -v ~/.cache/huggingface:/root/.cache/huggingface \
    --env "HF_TOKEN=<secret>" \
    --ipc=host \
    lmsysorg/sglang:latest \
    python3 -m sglang.launch_server \
        --model-path "iulio/FiscMind-Qwen38-27B-CoT" \
        --host 0.0.0.0 \
        --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "iulio/FiscMind-Qwen38-27B-CoT",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Docker Model Runner
How to use iulio/FiscMind-Qwen38-27B-CoT with Docker Model Runner:
```
docker model run hf.co/iulio/FiscMind-Qwen38-27B-CoT
```

🚀 Announcement: FiscMind-Qwen38-27B-CoT - Deliberative CoT Reasoning for Romanian Fiscal GAAP & Tax Law

by iulio - opened 1 day ago

Discussion

iulio

Owner 1 day ago

🇷🇴 Introducing FiscMind-Qwen38-27B-CoT

We are excited to announce the release of FiscMind-Qwen38-27B-CoT, a specialized 27-Billion parameter foundation-derived model fine-tuned for Romanian fiscal legislation, tax audit, and statutory double-entry accounting.

🌟 Key Highlights & Innovations

Cold-Start Deliberative Chain-of-Thought (<think>...</think>):
Instead of jumping directly to conclusion numbers, FiscMind structures its internal reasoning into a strict 4-phase audit trace:
- Legislație & Temei Legal: Citing exact articles from Codul Fiscal (Legea 227/2015), OMFP 1802/2014, Legea 296/2023, OUG 115/2023.
- Regim Fiscal & Deduceri: Resolving limitations (50% auto mixte, 33% plafon beneficii extrasalariale, scutiri VIES).
- Calcule Matematice: Exact taxable bases, VAT rates (19%, 9%, 5%), CAS (25%), CASS (10%), Impozit (10%), CAM (2.25%).
- SMT Formal Verification: Verifying that $\sum \text{Debit} = \sum \text{Credit}$ before generating final accounting entries.
Benchmark Results on FiscScore-54:
- Overall FiscScore: 91.2% (vs 78.4% baseline)
- CoT Reasoning Adherence: 98.0%
- Legal Citation Rate: 98.5%
- Double-Entry Balance (Debit = Credit): 94.4%
Roadmap to Etapa 2 (GRPO RLVR):
This cold-start adapter serves as the starting policy for Group Relative Policy Optimization (GRPO) with rule-based verifiers for mathematical balance ($R_{\text{SMT}}$), statutory citations ($R_{\text{Law}}$), and format guardrails ($R_{\text{Format}}$).

Feel free to test the model, review the model card, and provide feedback!

iulio

Owner 1 day ago

🎉 Model Training & Adapter Weights Successfully Deployed!

Architecture: Qwen 3.8-27B Dense CausalLM (Qwen/Qwen3.8-27B)
Stage: Cold-Start Deliberative Chain-of-Thought (CoT) SFT
Steps: 250 / 250 (Epoch 3.4)
Final Loss: 0.0938 (converged from 1.0909)
Artifacts: dapter_model.safetensors (159 MB), dapter_config.json, tokenizer & chat template
Verified Capabilities:
- Exact legal citation: Art. 273 Codul Fiscal (Legea 227/2015) for AIC, D390 VIES.
- Double-entry balance: Debit 371 (25.000 lei) == Credit 401 (25.000 lei), Autolichidare TVA 4426 = 4427 (4.750 lei).
- Perfect mathematical balance: 29.750 == 29.750 lei.

The model is now ready for Etapa 2: GRPO RLVR (Reinforcement Learning with Verifiable Rewards).

Upload images, audio, and videos by dragging in the text input, pasting, or clicking here.

Tap or paste here to upload images

· Sign up or log in to comment