mapomo-africa/models

Model cards, training and evaluation configurations. Weights live on Hugging Face # models **Model cards, training configurations and evaluation configurations.** Weights live on Hugging Face at huggingface.co, not in Git. This repository holds what a reader needs in order to judge a model: what it was trained on, what it is for, where it fails, and how it was measured. Status: **skeleton.** Cards land as models do. ## Why the cards are here and the weights are not Git is bad at large binaries, and Hugging Face is where the people who would evaluate an African-language speech model actually look. Separating them also means a model card can be corrected without moving gigabytes. ## Planned models | Model | Purpose | |---|---| | `mapomo-africa/outdoor-detect` | Detect and classify outdoor placements in field photographs | | `mapomo-africa/asr-wo` | Speech recognition, Wolof | | `mapomo-africa/asr-ff` | Speech recognition, Pulaar | | `mapomo-africa/asr-ha` | Speech recognition, Hausa | | `mapomo-africa/campaign-mention` | Identify campaign references in broadcast transcripts | ## Every card states its failure modes A card that lists only benchmark scores is not usable by someone deciding whether to trust an estimate built on it. Cards state where the model degrades: which accents, which recording conditions, which code-switching patterns, which lighting, which languages it was never trained on but will happily produce output for. This is not modesty. A speech model that silently fails on rural recordings produces a coverage gap that shows up as a campaign that appears to have spent nothing. ## Evaluation before deployment No model enters a deployment without an evaluation on data from the country and the languages of that deployment. A model evaluated only on read speech does not have a measured error rate on campaign rally audio, and treating one as the other is how a coverage gap becomes an apparent finding. ## License Apache-2.0 for configurations and cards. Individual model weights carry their own license on Hugging F …