dataeaze-RegLLM-zephyr_7b_beta-dzcompli

Model Card for Model ID

<!-- Provide a quick summary of what the model is/does. -->

RegLLM is LLM model for regulatory compliance. It has been domain adapted by unsupervised pretraining and instruction finetuned for regulatory compliance. This release focuses on Indian Banking rules and regulations.

Model Details

Model Description

<!-- Provide a longer summary of what this model is. -->

• Developed by: dataeaze systems pvt ltd
• Funded by: dataeaze systems pvt ltd
• Shared by: dataeaze systems pvt ltd
• Model type: MistralForCausalLM
• Language(s) (NLP): English
• License: cc-by-nc-sa-4.0 Model is made available under non-commercial use for research purposes only. For commercial usage please connect at [email protected]
• Finetuned from model: zephyr-7b-beta

Uses

<!-- Address questions around how the model is intended to be used, including the foreseeable users of the model and those affected by the model. -->

Direct Use

<!-- This section is for the model use without fine-tuning or plugging into a larger ecosystem/app. -->

The model has been crafted crafted to provide precise and insightful answers to a wide array of queries related to Indian Banking regulations.

Downstream Use

<!-- This section is for the model use when fine-tuned for a task, or when plugged into a larger ecosystem/app -->

This model can be used as core component in RegTech application

Out-of-Scope Use

<!-- This section addresses misuse, malicious use, and uses that the model will not work well for. -->

Model has been fine tuned on a specific task of answering questions related to Indian regulatory compliance. Any use beyond this is not guaranteed to be accurate.

Bias, Risks, and Limitations

<!-- This section is meant to convey both technical and sociotechnical limitations. -->

• Bias: Trained for English language only (as of now).
• Risk: Guardrails are reliant on the base models Mistral/Zephyr. Finetuning could impact this behaviour.
• Limitations: Intended to be a small model optimised for Indian regulations (as of now).

Recommendations

<!-- This section is meant to convey recommendations with respect to the bias, risk, and technical limitations. -->

• This model is supposed to be used as an assistive AI technology. Kindly consult and verify with the source documents for decision making.
• This model should be used with grounding on a set of regulatory documents.

How to Get Started with the Model

import torch
from transformers import pipeline

pipe = pipeline("text-generation", 
                model="dataeaze/dataeaze-RegLLM-zephyr_7b_beta-dzcompli", 
                torch_dtype=torch.bfloat16, 
                device_map="auto")

messages = [
    {
        "role": "system",
        "content": "You are a compliance assistant who answers in a formal manner",
    },
    {"role": "user", "content": "How often should IRRBB policies be reviewed?"},
]
prompt = pipe.tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
outputs = pipe(prompt, max_new_tokens=120, do_sample=True, temperature=0.1, top_k=50, top_p=0.95)
print(outputs[0]["generated_text"])

Sample Output

Question

How often should IRRBB policies be reviewed?

RegLLM respose

<|system|>
You are a compliance assistant who answers in a formal manner</s>
<|user|>
How often should IRRBB policies be reviewed?</s>
<|assistant|>
At least annually.
The review should be conducted by a committee consisting of the Chief Risk Officer, the Head of Market Risk, and the Head of Credit Risk.
The review should include an assessment of the effectiveness of the IRRBB policy and the adequacy of the capital and liquidity management framework.
The committee should also review the risk appetite and the risk tolerance of the organization.
The review should be documented and presented to the Board of Directors for approval.
The policy should be updated as necessary to reflect changes in the organization's business, products, or processes.

GPT-4 response

Reference

For evalating truthfulness / hallucination of this response, refer to RBI notification RBI/2022-23/180 DOR.MRG.REC.102/00-00-009/2022-23 (page 8)

Screenshot below

<img src="rbi_reference.png" alt="drawing" width="500"/>

As you can see, RegLLM has identified the frequency of IRRBB policies, while GPT-4 provides a more general response. Note, that the response of RegLLM is not backed by any external knowledge. When coupled with retriever model, RegLLM can provide fairly precise responses to user queries related to regulatory compliance.

Keep watching this space for more updates on the model and evaluations.

Model Card Authors

• Atharva Inamdar
• Niranjan Kakade
• Tony Tom
• Nayan Chheda
• Sourabh Daptardar

Model Card Contact

"dataeaze systems" [email protected]