Our Work

# LLMs Are Not a Silver Bullet: A Case Study on Software Fairness This repository contains the artifact for our empirical study on ML-based and LLM-based bias mitigation for tabular classification. …

# LLMs Are Not a Silver Bullet: A Case Study on Software Fairness This repository contains the artifact for our empirical study on ML-based and LLM-based bias mitigation for tabular classification. The current documentation follows the paper structure: - installation and experimental environment - repository structure for `ML/`, `LLM/`, `LLM/qwen_finetune/`, and `paper_results/` - reproduction commands for RQ1–RQ4 Detailed command examples are also provided in [docs/rq_commands.md](docs/rq_commands.md). ## Experimental Environment All experiments in the paper are implemented in Python 3.11.9. For the traditional ML paradigm, we use: - [IBM AIF360](https://github.com/Trusted-AI/AIF360) for bias mitigation methods and fairness metrics - `scikit-learn` for traditional classifiers - `TensorFlow Keras` for the DNN model For the LLM paradigm, we use API-based inference and fine-tuning: - OpenRouter for model inference - OpenAI API for fine-tuning The local Qwen fine-tuning implementation used in RQ4 is under `LLM/qwen_finetune/`. ## Installation We recommend using a virtual environment before installing the dependencies. ```bash python -m venv venv source venv/bin/activate pip install -r requirements.txt ``` If you encounter dependency issues related to AIF360, please refer to the official AIF360 installation page: - https://github.com/Trusted-AI/AIF360 ### API Keys Create a `.env` file in `LLM/`: ```bash OPENAI_API_KEY=your_openai_key OPENROUTER_API_KEY=your_openrouter_key ``` - `OPENAI_API_KEY` is required for GPT-based inference and OpenAI fine-tuning. - `OPENROUTER_API_KEY` is used when routing non-OpenAI models through OpenRouter. ## Repository Structure ### `ML/` This folder contains the traditional ML scripts used in RQ1. ML settings in the paper: - `default.py` - `cot.py` - `ltdd.py` - `fairmask.py` - `adv.py` - `maat.py` - `mirrorfair.py` - `eop.py` - `roc.py` Classifiers used in the paper: - `lr` - `svm` - `rf` - `dnn` ### `LLM/` This folder contains the main LLM pipeline used in RQ1–RQ4. Core scripts: - `main.py`: unified entry point - `zero_shot.py`: zero-shot inference - `few_shot.py`: few-shot inference - `finetune.py`: OpenAI fine-tuning pipeline LLM settings in the paper: - `zero_shot` - `few_shot` with `random` - `few_shot` with `label_flipping` - `few_shot` with `S1` (`--strategy balanced`) - `few_shot` with `S2` (`--strategy minority_balanced`) - `few_shot` with `S3` (`--strategy minority_unbalanced`) - `few_shot` with `FCG-S1` (`--strategy balanced --use_fcg`) - `few_shot` with `FCG-S2` (`--strategy minority_balanced --use_fcg`) - `few_shot` with `FCG-S3` (`--strategy minority_unbalanced --use_fcg`) - `finetune` with `random` - `finetune` with `cot` - `finetune` with `ltdd` - `finetune` with `fairmask`   ### `LLM/qwen_finetune/` This folder contains the local Qwen fine-tuning implementation used in RQ4. ### `paper_results/` This folder stores the reproducible paper results organized by research question: - `paper_results/rq1/` - `paper_results/rq2/` - `paper_results/rq3/` - `paper_results/rq4/` Paper-table summaries can be regenerated from these reproducible results with: ```bash cd Analysis_code python rq1.py python rq1_trans.py python rq2.py python rq3.py python rq4.py ``` Runtime outputs may also be generated under: - `ML/results/` - `LLM/results/` - `ML/predictions/` - `LLM/predictions/` ## Datasets The paper evaluates six dataset-attribute configurations derived from three benchmark datasets. | Name | Size | Sensitive attr(s) | Favorable label | Description | |------|------|-------------------|-----------------|-------------| | Adult | 48,843 | sex, race | income > 50K | Predict whether an individual's income exceeds 50K. | | COMPAS | 6,172 | sex, race | no recidivism | Predict criminal defendant recidivism. | | Credit | 30,000 | sex, age | default | Predict whether a customer will default on payment. | Sensitive attribute definitions: | Dataset | Attribute | Privileged group | Unprivileged group | |---------|-----------|------------------|--------------------| | Adult | sex | Male | Female | | Adult | race | White | Non-White | | COMPAS | sex | Female | Male | | COMPAS | race | Non-African-American | African-American | | Credit | sex | Female | Male | | Credit | age | `>=25 and <60` years | `<25 or >=60` years | ## Research Questions The reproduction commands for RQ1–RQ4 are provided in (docs/rq_commands.md).

Ushbu tezisda zamonaviy o‘zbek nasri va she’riyati Xurshid Do'stmuhammad va Usmon Azim misolida tinish belgilarining lingvopoetik imkoniyatlari hamda muallif punktuatsiyasining o‘zig…

Ushbu tezisda zamonaviy o‘zbek nasri va she’riyati Xurshid Do'stmuhammad va Usmon Azim misolida tinish belgilarining lingvopoetik imkoniyatlari hamda muallif punktuatsiyasining o‘ziga xos xususiyatlari tadqiq etilgan. Badiiy matnda tinish belgilarining faqat grammatik-sintaktik vosita emas, balki emotsional-ekspressivlik, psixologik taranglik va ritmik-intonatsion tuzilishni shakllantiruvchi polifunksional semiotik birlik ekanligi ilmiy dalillangan.

 The development of innovative management models in primary education institutions is becoming increasingly important in the context of digital transformation, changing educational needs, and the…

 The development of innovative management models in primary education institutions is becoming increasingly important in the context of digital transformation, changing educational needs, and the growing demand for learner-centred education. This article examines the theoretical and pedagogical foundations of developing an innovative management model for primary education institutions. The study focuses on the integration of strategic planning, digital technologies, data-driven decision-making, teacher collaboration, continuous monitoring, and stakeholder participation into school management. The proposed model combines administrative, pedagogical, technological, and analytical components and is designed to improve the quality of educational processes and institutional effectiveness. The research uses theoretical analysis, comparative analysis, system modelling, and generalisation of international educational research. The findings indicate that innovative management should move beyond traditional administrative functions toward a flexible, collaborative, evidence-based, and learner-centred approach. The proposed model provides a systematic framework for diagnosing institutional needs, planning development strategies, implementing innovations, monitoring outcomes, and continuously improving management practices.

Portable mixed-precision math, linear-algebra, & retrieval library with 2'000+ SIMD kernels for x86, Arm, RISC-V, LoongArch, Power, & WebAssembly, leveraging rare algebraic transforms with bot…

Portable mixed-precision math, linear-algebra, & retrieval library with 2'000+ SIMD kernels for x86, Arm, RISC-V, LoongArch, Power, & WebAssembly, leveraging rare algebraic transforms with both 1D & 2D registers like AMX & SME, covering 15+ numeric types from 4-bit integers & 6-bit floats to 128-bit complex numbers, validated against 118-bit extended-precision baselines with saturation, casting, & rounding edge-case coverage, in a 5-100x smaller binary than other BLAS-like alternatives, co-designed with Tensor abstractions in C++, Python, Rust, JavaScript, GoLang, & Swift.

This two minute video introduces the ARIADNE Research Infrastructure AISBL, a non-profit organisation, to researchers and end users of the Portal and its services. The 4 million records contained in t…

This two minute video introduces the ARIADNE Research Infrastructure AISBL, a non-profit organisation, to researchers and end users of the Portal and its services. The 4 million records contained in the Portal range from the earliest Hominids to the Cold War and cover four continents. The records to the whole archaeological process from excavation data, laboratory results through to publications and can be searched using a combination of "What", "Where" and "When".  Users can focus on the resources of interest and in many cases, directly download the information they need.

ForkUnion is a NUMA-aware fork-join thread-pool for systems languages, built for the tightest parallel-for loops rather than task queues. Its dispatch hot path makes no heap allocations, no system cal…

ForkUnion is a NUMA-aware fork-join thread-pool for systems languages, built for the tightest parallel-for loops rather than task queues. Its dispatch hot path makes no heap allocations, no system calls, and no compare-and-swap operations, and shares no cache line between workers, so fork-join latency stays flat as cores and sockets multiply. Idle workers park on a hardware address monitor where the processor offers one, waking the instant a peer writes the watched cache line, while per-domain memory placement and private per-worker cursors keep synchronization off the interconnect.

This record contains supplementary material for the extraction of entities and properties. The domain of this material is Industry 5.0. It documents both the prompt given to a Large Language Model (LL…

This record contains supplementary material for the extraction of entities and properties. The domain of this material is Industry 5.0. It documents both the prompt given to a Large Language Model (LLM) and the model-generated response. The prompt asks the LLM to extract for each competency question (CQ) the entities and properties that have to be introduced in the ontology. It requires the output for all CQs in a JSON format and gives a short example of that format. The response is a JSON array with 62 objects (one per CQ) that list the entities and properties needed to answer that question. The material is provided to support transparency, reproducibility and traceability of the LLM-assisted ontology engineering process.

This dataset holds maps of ice thickness and bed topography for all glaciers in the Randolf Glacier Inventory version 7.0.