dil-v1-insurance-model

DIL-V1 (Digital Insurance Language Model)

A fine-tuned LoRA adapter based on Qwen/Qwen2.5-3B-Instruct, trained for insurance domain reasoning, action routing, policy coverage retrieval, actuarial calculations, and regulatory compliance.

Key Capabilities

• Policy Coverage & Clause Retrieval: Evaluates claims against terms and exclusions.
• Actuarial Calculations: Pure premiums, loss ratios, combined ratios.
• Solvency II: Solvency Capital Requirement (SCR) and Minimum Capital Requirement (MCR).
• UK Insurance Regulations: FCA Consumer Duty, ORSA, and product governance.
• Multi-source Reasoning & Human Escalation: Identifies conflicting evidence and flags ambiguous cases for human review.

Quick Inference

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

BASE_MODEL = "Qwen/Qwen2.5-3B-Instruct"
ADAPTER = "Va1bhavdev/dil-v1-insurance-model"

tokenizer = AutoTokenizer.from_pretrained(ADAPTER)
base_model = AutoModelForCausalLM.from_pretrained(
    BASE_MODEL,
    torch_dtype=torch.bfloat16,
    device_map="auto"
)
model = PeftModel.from_pretrained(base_model, ADAPTER)

system_prompt = """You are DIL-V1, a Digital Insurance Language Model.

Your role is to understand the insurance scenario, identify which information
is relevant, prioritize the required evidence, select the appropriate capability,
and provide grounded reasoning.

Possible actions include:
- direct_answer
- retrieve
- tool
- solvency_tool
- currency_tool
- time_tool
- multi_step
- human_review

Do not invent policy wording, regulatory requirements, calculations or current
information. Use retrieval or tools when required."""

messages = [
    {"role": "system", "content": system_prompt},
    {"role": "user", "content": "What is an insurance excess?"}
]

prompt = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tokenizer(prompt, return_tensors="pt").to(model.device)

outputs = model.generate(**inputs, max_new_tokens=256)
print(tokenizer.decode(outputs[0][inputs.input_ids.shape[1]:], skip_special_tokens=True))