Nasseredd/datashi

DATASHI: A parallel English–Tashlhiyt corpus for orthography normalization and low-resource NLP. The dataset includes 5,000 sentence pairs and a 1,500-sentence expert-standardized subset for evaluation, along with scripts for normalization and analysis. # DATASHI: Orthography Normalization for Tashlhiyt Code and prompts for the paper: **DATASHI: A Parallel English–Tashlhiyt Corpus for Orthography Normalization and Low-Resource Language Processing** **Nasser-Eddine Monir, Zakaria Baou** --- ## Overview This repository contains the code, prompts, and evaluation setup used for orthography normalization experiments on Tashlhiyt with large language models. The task consists of mapping **non-standard Tashlhiyt (SHI-ns)** to **standardized Tashlhiyt (SHI-s)** under two prompting settings: - zero-shot - few-shot The evaluated models are: - GPT-5 - Claude Sonnet - Gemini 2.5 Pro - Mistral Large - Qwen3-Max --- ## Repository Structure ```text data/ inputs/ inputs.csv outputs/ claude/ gemini2.5/ gpt5/ mistral/ qwen3-max/ prompt/ few_shot_prompt.txt zero_shot_prompt.txt src/ evaluation.py run_inference.py .env README.md ``` --- ## Data The input CSV is structured as follows: ```text CATEGORY,EN,SHI-s,SHI-ns ``` Model inputs use the `SHI-ns` column. Each output file contains exactly one model inference per line, with no additional formatting. --- ## Run Inference Install dependencies: ```bash pip install openai anthropic google-generativeai mistralai python-dotenv ``` Create a `.env` file with the required API keys: ```bash OPENAI_API_KEY=... ANTHROPIC_API_KEY=... GOOGLE_API_KEY=... MISTRAL_API_KEY=... QWEN_API_KEY=... ``` Run inference: ```bash python src/run_inference.py ``` This produces: ```text data/outputs/ /zero_shot_normalization.txt data/outputs/ /few_shot_normalization.txt ``` --- ## Evaluation Run: ```bash python src/evaluation.py ``` The evaluation follows the paper setup and reports: * Word error rate (WER) * Levenshtein distance (LD) --- ## Dataset Note DATASHI contains 5,000 English–Tashlhiyt sentence pairs, including a 1,500-sentence subset with expert-standardized and non-standard versions for normalization experiments. The remaining non-standard Tashlhiyt sentences will …