| --- |
| language: |
| - en |
| license: cc-by-sa-4.0 |
| library_name: gguf |
| pipeline_tag: text-generation |
| base_model: Qwen/Qwen3.5-4B |
| base_model_relation: quantized |
| tags: |
| - ai-ethics |
| - organizational-ethics |
| - question-answering |
| - gguf |
| - qwen3.5 |
| - cpu-compatible |
| - local-inference |
| - faiss |
| - qdrant |
| - conversational |
| - knowledge-base |
| - arxiv |
| - governance |
| --- |
| |
| # AI Ethics Organizational Coach — Q&A and Advisory Model |
|
|
| **Demo Space:** *(coming soon)* |
| **Author:** [Roberto Lofaro](https://huggingface.co/robertolofaro) |
| **Bibliography Search:** [robertolofaro.com/searchkaggleaiethics_bibliography.php](https://robertolofaro.com/searchkaggleaiethics_bibliography.php) |
| **AI Ethics Primer Webapp:** [robertolofaro.com/aiethicsprimer](https://robertolofaro.com/searchkaggleaiethics.php) |
| **License:** [CC BY-SA 4.0](https://creativecommons.org/licenses/by-sa/4.0/) |
|
|
| --- |
|
|
| ## Model Overview |
|
|
| This is a **GGUF quantisation** of [Qwen/Qwen3.5-4B](https://huggingface.co/Qwen/Qwen3.5-4B), fine-tuned via a structured system prompt and optional retrieval layer to serve as an **AI Ethics organisational coach**: an expert consultant and philosopher focused on helping organisations assess the ethical impact of policy, organisational, and technological choices — specifically around introducing AI within organisational culture, systems, and processes. |
|
|
| The model's certified knowledge base is built from **959 ArXiv papers on AI Ethics**, curated monthly from the AI Ethics Primer project at [robertolofaro.com/aiethicsprimer](https://robertolofaro.com/searchkaggleaiethics.php). Selection criteria prioritise enabling communication on AI Ethics with both **technical and non-technical decision-makers**. The corpus has been updated monthly since August 2023; the HuggingFace model repository is updated on a **quarterly** basis. |
|
|
| --- |
|
|
| ## Intended Use |
|
|
| | Use | Supported | |
| |-----|-----------| |
| | Q&A on AI Ethics policies, frameworks, and practices | ✅ | |
| | Source recommendation from the ArXiv corpus | ✅ | |
| | Offline / local inference (CPU) | ✅ | |
| | General-purpose assistant | ⚠️ Not the primary intent | |
| | Definitive legal or compliance advice | ❌ (always complement with qualified advisors) | |
| | Commercial deployment without attribution | ❌ (see license) | |
|
|
| ### Primary Task |
|
|
| Given a natural language query from a decision-maker — technical or non-technical — the model delivers a structured advisory response grounded exclusively in its ArXiv corpus, following the three-part format described below. It bridges academic AI Ethics research and practical organisational guidance, making it suitable for governance teams, programme managers, AI strategy leads, and C-level executives preparing policy or adoption decisions. |
|
|
| --- |
|
|
| ## System Prompt |
|
|
| The model is configured with the following system prompt, which governs all interactions: |
|
|
| ``` |
| You are the "AI Ethics organizational coach," an expert consultant and philosopher |
| focused on helping organizations assessing the ethical impact of policy, organizational, |
| and technological choices, specifically introducing AI within organizational culture, |
| systems, processes. Your certified knowledgebase is represented by the 959 |
| ArXiv papers contained within the training database, selected to enabling communication |
| on AI Ethics with both technical and non-technical decision-makers. |
| |
| # Your Mission: |
| When a user asks a question, your goal is to provide a structured response based ONLY |
| on the ArXiv papers provided in your training. Do not provide general advice from |
| outside these sources. |
| |
| # Response Format: |
| 1. Executive Summary: A 2-3 sentence overview answering the core query. |
| 2. Guidelines & Hints: A markdown list of specific "answers/guidelines/hints" found |
| in the source material. |
| ``` |
|
|
| ### Sample Interaction |
|
|
| **Query:** |
| > *"What are the main risks of deploying AI in public-sector decision-making?"* |
|
|
| **Expected response structure:** |
|
|
| **Executive Summary:** |
| Based on the ArXiv corpus, the primary risks include algorithmic bias amplifying existing systemic inequalities, lack of transparency undermining accountability, and inadequate human oversight in high-stakes decisions. Several papers also flag procurement and governance gaps that allow under-regulated systems to enter public workflows. |
|
|
| **Guidelines & Hints:** |
| - **Algorithmic bias and fairness**: Multiple papers highlight how training data reflecting historical inequities can produce discriminatory outcomes in credit, hiring, and social benefit allocation. |
| - **Explainability requirements**: Papers on XAI (Explainable AI) emphasise that black-box models are inappropriate for decisions subject to legal challenge or democratic scrutiny. |
| - **Human-in-the-loop governance**: The corpus consistently recommends mandatory human review thresholds for consequential decisions, with clear escalation paths. |
| - **Procurement due diligence**: Several papers call for ethics impact assessments prior to public-sector AI procurement, analogous to environmental impact assessments. |
| - **Accountability gaps**: Where AI decisions cause harm, existing legal frameworks often leave affected citizens without clear redress mechanisms. |
|
|
| --- |
|
|
| ## About the Corpus |
|
|
| The 959 ArXiv papers span the following themes within AI Ethics: |
|
|
| - **Fairness, bias, and discrimination** in ML systems |
| - **Transparency, explainability, and accountability** (XAI, FATE) |
| - **AI governance, regulation, and policy** (EU AI Act, GDPR intersections) |
| - **Human-AI interaction** and organisational change management |
| - **AI safety and alignment** in deployed systems |
| - **Privacy and data rights** in AI pipelines |
| - **Societal and labour market impacts** of AI adoption |
| - **AI in high-stakes domains**: healthcare, public sector, finance, justice |
| - **Ethics of large language models** and generative AI |
|
|
| Papers are selected monthly from ArXiv and are searchable via the companion webapp at [robertolofaro.com/aiethicsprimer](https://robertolofaro.com/searchkaggleaiethics.php). Full bibliography browsable at [robertolofaro.com/searchkaggleaiethics_bibliography.php](https://robertolofaro.com/searchkaggleaiethics_bibliography.php). |
|
|
| **Update cadence:** corpus updated monthly; HuggingFace model repository updated quarterly. |
|
|
| --- |
|
|
| ## Available Quantisations |
|
|
| | Quantisation | File | Size | Recommended For | |
| |---|---|---|---| |
| | Q4\_K\_M | `aiethics-Q4_K_M.gguf` | ~2.71 GB | CPU inference, everyday use | |
| | Q8\_0 | `aiethics-Q8_0.gguf` | ~4.48 GB | Higher fidelity, 8 GB+ RAM | |
|
|
| The **Q4\_K\_M** variant is recommended for CPU-only environments. It is the default quantisation for Ollama and llama.cpp quick-start commands below. |
|
|
| --- |
|
|
| ## Usage |
|
|
| ### Quick Start with Ollama |
|
|
| ```bash |
| ollama run hf.co/robertolofaro/aiethics-model:Q4_K_M |
| ``` |
|
|
| ### Quick Start with llama.cpp |
|
|
| ```bash |
| # macOS / Linux |
| brew install llama.cpp |
| llama-server -hf robertolofaro/aiethics-model:Q4_K_M |
| |
| # Windows (WinGet) |
| winget install llama.cpp |
| llama-server -hf robertolofaro/aiethics-model:Q4_K_M |
| ``` |
|
|
| ### Quick Start with llama-cpp-python |
|
|
| ```python |
| from llama_cpp import Llama |
| |
| llm = Llama.from_pretrained( |
| repo_id="robertolofaro/aiethics-model", |
| filename="aiethics-Q4_K_M.gguf", |
| n_ctx=4096, |
| ) |
| |
| system_prompt = """You are the "AI Ethics organizational coach," an expert consultant |
| and philosopher focused on helping organizations assessing the ethical impact of policy, |
| organizational, and technological choices, specifically introducing AI within |
| organizational culture, systems, processes. Your certified knowledgebase is represented |
| by the 959 ArXiv papers contained within the training database, selected to |
| enabling communication on AI Ethics with both technical and non-technical |
| decision-makers. |
| |
| # Your Mission: |
| When a user asks a question, your goal is to provide a structured response based ONLY |
| on the ArXiv papers provided in your training. Do not provide general advice from |
| outside these sources. |
| |
| # Response Format: |
| 1. Executive Summary: A 2-3 sentence overview answering the core query. |
| 2. Guidelines & Hints: A markdown list of specific "answers/guidelines/hints" found |
| in the source material. |
| |
| response = llm.create_chat_completion( |
| messages=[ |
| {"role": "system", "content": system_prompt}, |
| { |
| "role": "user", |
| "content": "What frameworks exist for AI ethics auditing in enterprises?" |
| } |
| ] |
| ) |
| print(response["choices"][0]["message"]["content"]) |
| ``` |
|
|
| ### Quick Start with Docker |
|
|
| ```bash |
| docker model run hf.co/robertolofaro/aiethics-model:Q4_K_M |
| ``` |
|
|
| --- |
|
|
| ## Retrieval-Augmented Variants |
|
|
| The repository includes reference implementations demonstrating different retrieval strategies. The system prompt alone yields well-grounded advisory responses; embedding-based variants add precision for longer, more ambiguous, or cross-domain queries. |
|
|
| ### Mode A — System Prompt Only (no embeddings) |
|
|
| Fastest option. Relies entirely on the structured system prompt encoding the corpus themes. No vector index required; runs on any machine with llama-cpp-python installed. |
|
|
| ```bash |
| python samples_hf/run_no_embeddings.py \ |
| --query "How should organisations govern AI procurement decisions?" |
| ``` |
|
|
| ### Mode B — FAISS-HNSW Index |
|
|
| Uses a pre-built FAISS index (HNSW graph) over sentence-transformer embeddings of paper abstracts and key passages. Suitable for environments where FAISS is available and a persistent index is desirable. |
|
|
| ```bash |
| # First-run: builds the index (saved locally) |
| python samples_hf/run_faiss_hnsw.py --build-index |
| |
| # Subsequent runs: load existing index |
| python samples_hf/run_faiss_hnsw.py \ |
| --query "Bias mitigation in automated hiring systems" |
| ``` |
|
|
| ### Mode C — Qdrant Vector Store |
|
|
| Uses a local Qdrant instance (or Qdrant Cloud) as the vector store. Preferred for production-style deployments or when persistence, filtering by paper metadata, and collection management are required. |
|
|
| ```bash |
| # Start Qdrant locally (Docker) |
| docker run -p 6333:6333 qdrant/qdrant |
| |
| # Upsert embeddings and query |
| python samples_hf/run_qdrant.py \ |
| --query "Accountability gaps in public-sector AI deployment" |
| ``` |
|
|
| --- |
|
|
| ## System Prompt Design |
|
|
| The system prompt is the primary configuration layer of the model. It: |
|
|
| - Establishes the **AI Ethics organisational coach** persona, positioned as expert consultant and philosopher |
| - Scopes responses **exclusively** to the ArXiv corpus — no out-of-corpus general advice |
|
|
| --- |
|
|
| ## Companion Webapp |
|
|
| An interactive bibliography search interface is available at: |
|
|
| 🔗 **[robertolofaro.com/aiethicsprimer](https://robertolofaro.com/aiethicsprimer)** |
|
|
| This webapp enables tag-based and keyword search across the full corpus of 959 papers, and serves as a complement to model-generated recommendations. The full bibliography is browsable at: |
|
|
| 🔗 **[robertolofaro.com/searchkaggleaiethics_bibliography.php](https://robertolofaro.com/searchkaggleaiethics_bibliography.php)** |
|
|
| A Gradio-based interactive demo Space is planned; it will run the Q4\_K\_M quantisation on CPU hardware. Announcement will be made via [Linkedin](https://linkedin.com/in/robertolofaro) and [Patreon](https://patreon.com/robertolofaro). |
|
|
| --- |
|
|
| ## Limitations |
|
|
| - Recommendations are bounded by the 959 ArXiv papers in the corpus at training time; the model will not draw on sources outside this set. |
| - The model does not have live internet access; content reflects the corpus as indexed at the last quarterly build. |
| - Papers added in the most recent monthly update batch may not be reflected until the next quarterly HuggingFace release. |
| - CPU inference with Q4\_K\_M typically yields response times of 15–60 seconds depending on hardware; Q8\_0 benefits from GPU acceleration; adjust the ctx as needed. |
| - The model is advisory in nature; outputs should be treated as structured research summaries, not as legal, compliance, or regulatory advice. Always complement with qualified professional guidance for consequential decisions. |
| - Due to its content (many papers share similar or overlapping material), the answers up to the </think> could be prone to hallucinations and repetitions. |
| - The model inherits any biases present in the Qwen3.5-4B base model; standard critical judgement should be applied to outputs. |
| |
| --- |
| |
| ## Ethical Considerations |
| |
| - The corpus consists entirely of open-access ArXiv papers; no third-party paywalled content is embedded. |
| - The advisory system is informational and does not collect user data. |
| - The model is explicitly designed to support human oversight rather than replace it — consistent with the AI governance principles it advises on. |
| - Users in regulated industries (finance, healthcare, public sector) should treat model outputs as a research starting point, not as compliance guidance. |
| - The model inherits any selection biases present in the curation process; the monthly update cycle and open bibliography search aim to maintain transparency about corpus composition. |
| |
| --- |
| |
| ## Citation & DOI |
| |
| Model DOI: **[10.57967/hf/8841](https://doi.org/10.57967/hf/8841)** |
| |
| ```bibtex |
| @misc{lofaro2026aiethicsmodel, |
| author = {Roberto Lofaro}, |
| title = {AI Ethics Organizational Coach — Q\&A and Advisory Model}, |
| year = {2026}, |
| doi = { 10.57967/hf/8841 }, |
| url = {https://huggingface.co/robertolofaro/aiethics-model}, |
| note = {GGUF quantisation of Qwen3.5-4B, fine-tuned for AI Ethics advisory |
| via structured system prompt and optional retrieval (FAISS-HNSW / |
| Qdrant); corpus of 959 ArXiv papers on AI Ethics, updated quarterly} |
| } |
| ``` |
| |
| --- |
| |
| ## License |
| |
| This model card and associated scripts are released under **[CC BY-SA 4.0](https://creativecommons.org/licenses/by-sa/4.0/)**. |
| The base model weights are subject to the [Qwen3 License](https://huggingface.co/Qwen/Qwen3-4B/blob/main/LICENSE). |
| |
| --- |
| |
| *Published openly as part of Roberto Lofaro's AI-assisted knowledge production initiative.* |
| *[GitHub](https://github.com/robertolofaro) · [Patreon](https://patreon.com/robertolofaro) · [robertolofaro.com](https://robertolofaro.com)* |
| |
