Mission

ZamAI builds practical AI resources for Pashto, Dari, and Afghan-language workflows: datasets, model checkpoints, evaluation assets, and usable demos. The goal is to make Afghan-language AI easier to inspect, load, improve, and deploy.

Canonical Collections

Use these collections as the clean front door. They separate primary resources from experiments and older work.

• ZamAI Pashto Text Datasets
• ZamAI Pashto Speech Datasets
• ZamAI Pashto Vision OCR Datasets
• ZamAI Pashto Models

Primary Datasets

Model Status

The model collection separates loadable checkpoints from metadata-only or incomplete checkpoint repos.

Loadable / usable checkpoints

• tasal9/pashto-base-bloom
• tasal9/ZamAI-LIama3-Pashto
• tasal9/Multilingual-ZamAI-Embeddings
• tasal9/ZamAI-Phi-3-Mini-Pashto
• tasal9/ZamAI-Sentiment-Pashto
• tasal9/ZamAI-mT5-Pashto

Experimental / incomplete

These repos are kept public for transparency, planning, and reproducibility notes, but their cards clearly state if weights or tokenizer files are missing:

• tasal9/ZamAI-Mistral-7B-Pashto
• tasal9/zamai-qwen2.5-7B-instruction
• tasal9/ZamAI-Whisper-v3-Pashto
• tasal9/ZamAI-Pashto-Translator-FacebookNLB-ps-en
• tasal9/ZamAI-QA-Pashto

Spaces

The public Spaces are demos and workbenches for Pashto AI workflows. Current reviewed Spaces have valid cards; build issues are fixed where possible. Some Spaces may sleep or pause to control compute cost.

Contribution Priorities

The highest-impact work right now:

1. Correct and validate Pashto speech transcripts.
2. Upload missing model weight/tokenizer artifacts for incomplete repos.
3. Add lightweight evaluations for primary models.
4. Keep canonical collections up to date as repos mature.

Contact

• Website: https://zamai.dev
• Hugging Face: https://huggingface.co/ZamAI-Pashto
• Founder: Yaqoob Tasal