Weekly Huggingface Model Digest

Weekly Hugging Face Model Digest

Workflow

1. Read STATE.md; for a new execution, reset it from references/template/STATE.template.md. Read references/cron.md for scheduled-task metadata.
2. Resolve the current month and timestamp in Asia/Taipei. Inspect the newest existing report under ~/Desktop/daily/huggingface-model-survey/ and retain its heading and field structure when it is readable.
3. Query current Hugging Face model metadata and keep only models whose createdAt falls in the current month. Group by text-generation, text-to-image, and text-to-audio.
4. Sort each category by likes descending and retain at most five models. Do not replace this with a download-count or all-time popularity ranking.
5. For each model, record publication time, likes, model size (or 未知), key features, training dataset when available, and the direct Hugging Face model URL. If a model is explicitly marked uncensored, state 移除 safety 限制 and report whether evidence confirms a claude-4.6-opus-dataset fine-tune.
6. Write the report to ~/Desktop/daily/huggingface-model-survey/YYYY-MM-DD_HH-MM.md, creating the parent directory if needed.
7. Verify the file and update STATE.md with the query filters, category counts, and saved path.

Rules

• Use current, source-backed metadata; do not infer an uncensored label, dataset, size, or publication date without evidence.
• If a field is unavailable, write 未知 instead of fabricating it.
• Use ### 1) 模型名 style headings and bulleted model fields.
• If writing fails, report the exact error and a concrete retry suggestion.

Output

Report the full saved path and briefly note any category with fewer than five qualifying models.