This model is a merged, domain-adapted version of Meta-Llama-3.1-8B-Instruct, fine-tuned for regulatory question answering in engineering domains, specifically ABS (American Bureau of Shipping) regulations.

The model is designed to generate structured, context-aware answers grounded in regulatory text and is intended for use in compliance and engineering reasoning tasks.

Base Model

meta-llama/Meta-Llama-3.1-8B-Instruct

Model Type

This repository contains a fully merged model:

LoRA adapters were trained using Unsloth
The adapters were merged into the base model using merge_and_unload()
The released checkpoint is standalone and does not require PEFT

Training Details

Fine-tuning method: LoRA (parameter-efficient fine-tuning)
Framework: Unsloth
Optimizer: AdamW

Dataset

The model was trained on domain-specific question–answer pairs derived from ABS regulatory documents.

Data format:

JSONL with chat-style messages
Prompt: all prior messages
Target: final assistant response

Intended Use

Regulatory question answering (ABS rules)
Engineering compliance support
Domain-specific reasoning
Integration into RAG (Retrieval-Augmented Generation) systems

Limitations

May produce incorrect numerical or threshold-based reasoning
Performance depends on input prompt and external context
Not suitable for safety-critical or legal decisions without expert validation
Domain specialization may reduce general-purpose performance

Usage

python

from transformers import AutoModelForCausalLM, AutoTokenizer

repo_id = "Sara121/Llama-3.1-8B-ABS-Regulatory"

tokenizer = AutoTokenizer.from_pretrained(repo_id)
model = AutoModelForCausalLM.from_pretrained(
    repo_id,
    torch_dtype="auto",
    device_map="auto",
)

model.eval()

Sara121/Llama_3.1_8B_ABS_Regulatory

Llama-3.1-8B-ABS-Regulatory

Overview