morbi-v022-lora

<div align="center">

Morbi v0.2.2

Professional Domain Expert for Insurance, Actuarial Science, Accounting & Legal

A specialized AI assistant trained on 19,000+ examples across insurance, actuarial science, accounting (CPA/CFA), and legal (Bar Exam) domains.

HEIR Platform | Documentation | API Access

</div>

---

Model Description

Morbi is a domain-specialized large language model built for professional financial services applications. Named after the actuarial term for mortality/morbidity analysis, Morbi provides expert-level assistance across interconnected professional domains.

What's New in v0.2.2

• Accounting Domain: Added CPA exam topics (FAR, AUD, REG, BEC), GAAP/IFRS standards, and financial calculations
• Legal Domain: Added California Bar Exam subjects, UCC Article 2, and case analysis capabilities
• Improved Training: Lower learning rate with cosine schedule for better convergence

---

Intended Uses

Primary Use Cases

Out-of-Scope Uses

• Not for: Medical diagnosis, investment advice requiring fiduciary duty, legal representation
• Not a substitute for: Licensed professionals (CPAs, attorneys, actuaries, financial advisors)
• Should not be used for: High-stakes decisions without professional verification

---

Domain Expertise

Insurance & Actuarial Science

Life Insurance          Health Insurance         Actuarial Methods
├── Term Life           ├── Medical              ├── Mortality Tables (CSO, VBT)
├── Whole Life          ├── Dental/Vision        ├── Interest Theory
├── Universal Life      ├── Disability (STD/LTD) ├── Life Contingencies
├── Variable Life       ├── Long-Term Care       ├── Reserve Calculations
├── Indexed Products    └── Medicare/Medicaid    ├── Premium Development
└── Annuities                                    └── Experience Studies

Sample Topics: Non-forfeiture options, policy loans, 1035 exchanges, HIPAA compliance, NAIC regulations, SOA exam concepts, risk classification, reinsurance structures.

Accounting & Finance (CPA/CFA)

Legal (Bar Exam)

---

Training Details

Training Data

Data Sources:

• Proprietary insurance training materials
• Actuarial exam preparation content
• CPA/CFA study guides and practice questions
• California Bar Exam past papers and study materials
• UCC Article 2 and Black's Law Dictionary excerpts

Training Procedure

LoRA Configuration

Quantization

Hardware & Compute

---

Evaluation

Quantitative Results

Qualitative Examples

<details> <summary><b>Insurance Example</b></summary>

Prompt: What are the key differences between term and whole life insurance?

Response: Term life insurance provides pure death benefit protection for a specified period (10, 20, or 30 years) with no cash value accumulation. Premiums are lower initially but increase upon renewal. Whole life insurance provides lifetime coverage with guaranteed level premiums and builds cash value on a tax-deferred basis. The cash value can be accessed through policy loans or withdrawals. Key considerations include: (1) affordability vs. permanence, (2) need for cash value accumulation, (3) estate planning objectives, and (4) the insured's age and health status...

</details> <details> <summary><b>Actuarial Example</b></summary>

Prompt: Calculate the present value of a life annuity-due for a person age 65.

Response: To calculate the present value of a life annuity-due (ä_65), we use the formula: ä_x = Σ(v^k × k_p_x) for k=0 to ω-x, where v = 1/(1+i) is the discount factor, k_p_x is the probability of surviving k years, and ω is the limiting age. Using the 2017 CSO mortality table at 4% interest...

</details> <details> <summary><b>Accounting Example</b></summary>

Prompt: When does UCC Article 2 apply instead of common law for contracts?

Response: UCC Article 2 applies to contracts for the sale of goods, while common law governs contracts for services, real estate, and intangibles. Key distinctions: (1) Goods are defined as movable, tangible personal property at the time of identification to the contract; (2) Mixed contracts (goods + services) use the predominant purpose test; (3) UCC provides different rules for offer/acceptance (battle of forms §2-207), statute of frauds ($500 threshold), and warranties...

</details> <details> <summary><b>Legal Example</b></summary>

Prompt: Explain the hearsay rule and its major exceptions.

Response: Hearsay is an out-of-court statement offered to prove the truth of the matter asserted (FRE 801). It is generally inadmissible because the declarant was not under oath and cannot be cross-examined. Major exceptions include: (1) Present sense impression (FRE 803(1)); (2) Excited utterance (FRE 803(2)); (3) State of mind (FRE 803(3)); (4) Medical diagnosis (FRE 803(4)); (5) Business records (FRE 803(6)); (6) Former testimony (FRE 804(b)(1)); (7) Dying declaration (FRE 804(b)(2))...

</details>

---

Usage

Quick Start

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

# Load base model
base_model = AutoModelForCausalLM.from_pretrained(
    "mistralai/Mistral-Small-Instruct-2409",
    torch_dtype=torch.bfloat16,
    device_map="auto",
    trust_remote_code=True
)

# Load LoRA adapter
model = PeftModel.from_pretrained(base_model, "h3ir/morbi-v022-lora")
tokenizer = AutoTokenizer.from_pretrained("mistralai/Mistral-Small-Instruct-2409")

# Generate response
prompt = """<s>[INST] You are Morbi, an expert AI assistant specializing in insurance, actuarial science, accounting, and legal matters.

What is the difference between GAAP and IFRS for revenue recognition? [/INST]"""

inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
outputs = model.generate(
    **inputs,
    max_new_tokens=512,
    temperature=0.7,
    do_sample=True,
    top_p=0.9
)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

Prompt Format

Morbi uses the Mistral instruction format with a specialized system prompt:

<s>[INST] You are Morbi, an expert AI assistant specializing in insurance, actuarial science, accounting, and legal matters.

{your question here} [/INST]

With vLLM

from vllm import LLM, SamplingParams
from vllm.lora.request import LoRARequest

llm = LLM(
    model="mistralai/Mistral-Small-Instruct-2409",
    enable_lora=True,
    max_lora_rank=32
)

sampling_params = SamplingParams(temperature=0.7, max_tokens=512)

outputs = llm.generate(
    prompts,
    sampling_params,
    lora_request=LoRARequest("morbi", 1, "h3ir/morbi-v022-lora")
)

With Text Generation Inference (TGI)

# Download and merge adapter first, or use TGI's LoRA support
docker run --gpus all -p 8080:80 \
  -v $PWD/data:/data \
  ghcr.io/huggingface/text-generation-inference:latest \
  --model-id mistralai/Mistral-Small-Instruct-2409 \
  --lora-adapters h3ir/morbi-v022-lora

---

Technical Specifications

Memory Requirements

Supported Backends

• Transformers + PEFT
• vLLM (with LoRA support)
• Text Generation Inference
• llama.cpp (after conversion)
• Ollama (after conversion)

---

Limitations & Risks

Known Limitations

1. Domain Imbalance: Insurance/actuarial data (98.5%) significantly outweighs accounting (0.8%) and legal (0.7%) data
2. Jurisdiction: Legal training focuses on U.S. law (California Bar Exam); may not apply to other jurisdictions
3. Temporal: Training data has a knowledge cutoff; regulations and laws change
4. Calculations: While trained on calculation examples, complex numerical computations should be verified

Potential Biases

• May reflect biases present in insurance industry practices
• Legal analysis may favor common law interpretations over civil law systems
• Financial calculations assume U.S. regulatory frameworks

Mitigation Recommendations

• Always verify critical calculations with qualified professionals
• Cross-reference legal advice with licensed attorneys
• Use for educational and research purposes; not as sole decision-making tool

---

Environmental Impact

*Estimate based on U.S. average grid intensity

---

Version History

---

Citation

@misc{morbi-v022,
  author = {HEIR},
  title = {Morbi v0.2.2: Professional Domain Expert for Insurance, Actuarial Science, Accounting \& Legal},
  year = {2026},
  month = {January},
  publisher = {Hugging Face},
  howpublished = {\url{https://huggingface.co/h3ir/morbi-v022-lora}},
  note = {LoRA adapter for Mistral-Small-Instruct-2409}
}

---

Contact & Support

• Website: https://heir.es
• Documentation: https://heir.es/docs
• Issues: GitHub Issues
• Email: [email protected]

---

<div align="center">

Built with care by HEIR

Empowering wealth management through AI

</div>