llama-scheme-a

LLM-VNSE: Llama 3.1 8B Fine-tuned for QoS-Aware Volunteer Node Selection

Model Description

This repository contains four QLoRA fine-tuned adapters for Llama 3.1 8B Instruct, one for each QoS weight scheme (A, B, C, D), trained as part of the LLM-VNSE framework for QoS-aware volunteer node selection in fog computing environments. The adapters are released as reproducibility artefacts accompanying the paper:

"Calibrated and Adaptive Listwise Ranking with Large Language Models for Quality-of-Service-Aware Volunteer Node Selection in Fog Computing"

Available Adapters

Training Data

The adapters were fine-tuned on the QWS Dataset 2.0 (2,507 real web services across five QoS attributes: response time, throughput, availability, reliability, and latency). The dataset was partitioned into 2,005 training lists, 251 validation lists, and 251 test lists following an 80/10/10 ratio. Ground-truth rankings were generated using a weighted-sum scoring function under each scheme's weight configuration as defined in Table 1 of the paper.

Intended Use

These adapters are intended for QoS-aware listwise ranking of volunteer nodes in fog computing environments. Given a candidate pool of volunteer nodes with their QoS profiles, the fine-tuned model produces a complete ranked ordering from best to worst under the active SLA priority configuration.

Training Configuration

• Method: QLoRA (4-bit NF4 quantisation)
• LoRA Rank: 16
• LoRA Alpha: 32
• Learning Rate: 2 × 10⁻⁴
• Epochs: 3
• Effective Batch Size: 32
• Hardware: NVIDIA A100 40GB GPU (Google Cloud Compute Engine)
• Random Seed: 42

Known Limitations

1. Dataset distribution: The model was trained on QWS Dataset 2.0 collected in 2008, which describes stable server-hosted web services. Transferability to real volunteer fog node deployments may be limited due to distributional differences in response time and throughput.

2. ICL instability: When used as a base model under in-context learning without fine-tuning, Llama 3.1 8B exhibits severe confidence compression (scores clustering around 0.5) and sensitivity to prompt serialisation format, with accuracy drops of 17–23% under Mathematical Vector/Tuple Notation compared to the original key-value format. Fine-tuning with these adapters overcomes these limitations.

3. Scheme-specific: Each adapter is optimised for its corresponding weight configuration. For cross-scheme generalisation without retraining, use the weight-conditioned Protocol II approach described in the paper.

Citation

@article{alsaiari-llmvnse,
  title={Calibrated and Adaptive Listwise Ranking with Large Language Models for Quality-of-Service-Aware Volunteer Node Selection in Fog Computing},
  author={Alsaiari, Najat and Hussain, Farookh},
  year={2026}
}

Repository

Code and reproducibility artefacts: github.com/NajatAlsaiari/LLM-VNSE