nirasha/KGFactExplainer-Qwen3-8B-LoRA

KGFactExplainer: Qwen3-8B LoRA Adapter

This repository contains the LoRA adapter weights for Qwen3-8B fine-tuned for the evidence-path generation component of KGFactExplainer, a knowledge graph-based approach for explaining fact inclusion in abstractive summarization.

The adapter is designed to generate traversal-compatible evidence paths connecting summary facts to information in the source-document knowledge graph.

Model Details

Model Description

KGFactExplainer uses a knowledge graph to represent information in a source document and generates evidence paths that explain how facts appearing in an abstractive summary are supported by the source.

This adapter was obtained by parameter-efficient fine-tuning of Qwen3-8B using Low-Rank Adaptation (LoRA). The training data were generated using a larger teacher model and used to transfer evidence-path generation behaviour to the smaller student model.

The adapter is intended to be used together with the corresponding Qwen3-8B base model and the KGFactExplainer evidence-path generation pipeline.

• Model type: Large language model LoRA adapter
• Base model: unsloth/qwen3-8b-unsloth-bnb-4bit
• Fine-tuning method: LoRA / PEFT
• Task: Knowledge-graph evidence-path generation
• Application: Fact inclusion explanation for abstractive summarization
• Primary language: English
• Framework: Hugging Face Transformers, PEFT, TRL, Unsloth

Model Sources

• Base model: unsloth/qwen3-8b-unsloth-bnb-4bit
• Project repository: [Add GitHub repository URL]
• Research paper: [Add paper URL when available]

Intended Use

Direct Use

The adapter is intended for research use in generating evidence paths over source-document knowledge graphs.

A typical input consists of a summary fact represented as a subject-predicate-object (SPO) triplet together with the relevant knowledge-graph context. The model generates a path intended to connect the summary fact to supporting information in the source document.

Downstream Use

The adapter can be integrated into systems for:

• source-grounded summarization analysis;
• fact verification;
• knowledge-graph-based explanation;
• explainable abstractive summarization;
• evidence retrieval and evidence-path generation.

The adapter should be used as part of the KGFactExplainer pipeline rather than interpreted as a general-purpose summarization or question-answering model.

Out-of-Scope Use

This adapter was not developed or evaluated as a general-purpose conversational assistant, general-purpose summarizer, or general-purpose fact-checking model.

It should not be assumed to provide reliable factual verification outside the knowledge-graph and source-grounded setting for which it was trained and evaluated.

Limitations

The model has several limitations:

• It was trained and evaluated within the KGFactExplainer experimental setting.
• Its evidence-path generation quality depends on the quality and coverage of the underlying source-document knowledge graph.
• Generated paths may contain unsupported or incorrectly connected relations.
• Performance outside the datasets, domains, and graph construction procedures used in the associated research has not been established.
• The adapter is not intended to replace independent verification of factual claims.
• As a LoRA adapter, it requires the corresponding compatible base model and PEFT-compatible inference setup.

How to Get Started

The adapter can be loaded using Hugging Face Transformers and PEFT.

from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

base_model = "unsloth/qwen3-8b-unsloth-bnb-4bit"
adapter_model = "[YOUR_HUGGINGFACE_USERNAME]/[YOUR_ADAPTER_REPOSITORY]"

tokenizer = AutoTokenizer.from_pretrained(base_model)

model = AutoModelForCausalLM.from_pretrained(
    base_model,
    device_map="auto"
)

model = PeftModel.from_pretrained(
    model,
    adapter_model
)

The exact inference configuration should follow the configuration used in the KGFactExplainer experiments.

Training Details

Training Data

The adapter was trained using a generated dataset constructed for the KGFactExplainer evidence-path generation task.

The training examples were generated using a larger teacher language model and contain knowledge-graph evidence-path supervision. The generated data were designed to teach the student model to produce traversal-compatible paths connecting summary facts to supporting source-document information.

The underlying source documents are derived from the XSum dataset. Users reproducing the training data should obtain the original dataset from its official source and follow its applicable licensing and usage conditions.

[Add link to the generated training dataset here, if publicly released.]

Training Procedure

The student model was fine-tuned using Low-Rank Adaptation (LoRA) with parameter-efficient fine-tuning.

The resulting adapter contains the learned LoRA parameters rather than a full copy of the base model.

Training Hyperparameters

The complete training configuration is provided in the associated project repository.

Key configuration details:

• Base model: Qwen3-8B
• Fine-tuning method: LoRA
• Training framework: Hugging Face Transformers / PEFT / TRL / Unsloth
• Precision: 4-bit quantized base model during training
• Task: Evidence-path generation

[Add LoRA rank, alpha, dropout, learning rate, batch size, number of epochs/steps, and other values if you want the Hugging Face artifact to be fully reproducible.]

Evaluation

The adapter was evaluated as part of the KGFactExplainer experimental pipeline.

Evaluation focuses on the quality of generated evidence paths, including agreement between generated paths and reference evidence paths.

The associated research paper reports the experimental methodology, datasets, evaluation metrics, and results.

Technical Specifications

Model Architecture

The adapter is based on Qwen3-8B and modifies the base model through LoRA parameter updates.

The adapter itself does not contain the complete base model weights.

Compute Infrastructure

Training was performed in a Google Colab environment using an NVIDIA L4 GPU. The Qwen3-8B base model was loaded in 4-bit quantized form using Unsloth, and LoRA-based parameter-efficient fine-tuning was used to reduce GPU memory requirements.

Acknowledgements

This artifact was developed as part of research on knowledge-graph-based explanations for abstractive summarization.

License

[Add the applicable license here.]

Contact

[Add project/research contact information.]