paulinechaga/Safety_LLM_Challenge

Can Africa’s AI withstand its own linguistic and cultural complexity? # Safety_LLM_Challenge Submission repository for the **Zindi African Trust & Safety LLM Challenge** — a red‑teaming competition that asks participants to surface trust and safety weaknesses in African‑trained / Africa‑deployed Large Language Models. This repository targets the **Swahili** track, against the model `sartifyllc/Pawa-Gemma-Swahili-2B`. --- ## Purpose and ethical framing The goal of the challenge — and of this repository — is **defensive**: to systematically document where a publicly deployed Swahili LLM fails on safety, so that model providers, auditors, and downstream deployers in African markets can close those gaps before the model reaches real users (banking, healthcare, telecoms, education). The adversarial prompts in this repository are submitted to a sanctioned competition under a CC‑BY‑SA 4.0 licence, as required by Zindi. They are **not intended for use against live production systems**, for harming users, or for bypassing safety controls outside the controlled evaluation context. If you are reading this repository outside the Zindi evaluation, please treat the contents as a safety audit artefact, not as a toolkit. --- ## Repository contents | File | What it is | |---|---| | `SampleSubmission.md` | Official Zindi template showing the required structure for each attack entry. | | `attack_types.csv` | Official taxonomy of attack strategies (e.g. `roleplay`, `prompt_injection`, `context_poisoning`). | | `risk_categories.csv` | Top‑level risk categories (e.g. `cybersecurity`, `unsafe_medical_advice`). | | `risk_subcategories.csv` | Valid (category, subcategory) pairs used for labelling. | | `manifest-…json` | Zindi‑provided challenge manifest. | | `attacks.py` | Python list of the Swahili adversarial prompts used for this submission, with metadata, English translations, and notes. | | `run_attacks.py` | Harness that loads the target model, runs each prompt, and writes the submission markdown + raw JSON. | | `raw_responses.json` | Raw mode …