AI4SciNoob/AskCGSim-Adapters

AskCGSim Adapters

LoRA/QLoRA adapters used from the CGSim project.

These adapters correspond to the model variants evaluated in the upcoming paper:

CGSim: A Simulation Framework for Large Scale Distributed Computing Infrastructures with AI-Driven Applications

Base Model

NVIDIA/Llama-3.1-Nemotron-Nano-8B-v1

Repository Structure

AskCGSim-Adapters/
├── CoT/
├── CoT_WeightedLoss/
├── NoCoT/
└── OnlySQL/

Adapter Variants

• CoT: reasoning + database inspection tool calling
• CoT_WeightedLoss: CoT model trained with weighted cross-entropy loss
• NoCoT: database inspection tool calling without explicit reasoning traces
• OnlySQL: direct SQL generation without reasoning or tool calling

Training Configuration

The models were fine-tuned using QLoRA with:

• LoRA rank: 16
• LoRA alpha: 32
• LoRA dropout: 0.05
• Target modules: q, k, v, o
• Batch size: 1
• Gradient accumulation steps: 8

The training follows a two-stage procedure:

• Stage 1: 500 template-generated samples, learning rate 1e-4
• Stage 2: 898 human-reviewed samples, learning rate 5e-5

The CoT_WeightedLoss variant additionally uses weighted cross-entropy loss for SQL- and JSON-sensitive tokens.

Usage

Install the required packages:

pip install transformers peft accelerate bitsandbytes

Load the base model and adapter:

from transformers import AutoModelForCausalLM
from peft import PeftModel

base_model = AutoModelForCausalLM.from_pretrained(
    "NVIDIA/Llama-3.1-Nemotron-Nano-8B-v1"
)

model = PeftModel.from_pretrained(
    base_model,
    "AI4SciNoob/AskCGSim-Adapters",
    subfolder="CoT/checkpoint-210"
)

Checkpoint Contents

Each directory contains the corresponding full training checkpoint, including files such as:

adapter_config.json
adapter_model.safetensors
chat_template.jinja
tokenizer.json
tokenizer_config.json
trainer_state.json
optimizer.pt
scheduler.pt
rng_state.pth

Paper

This repository accompanies the upcoming paper:

CGSim: A Simulation Framework for Large Scale Distributed Computing Infrastructures with AI-Driven Applications

Citation information will be added after publication.