Spaces:

---

huggingface

/

hub-stats

Like 68

Running

Adds the State of OSS (Oct 2026) stats that can be computed from hub-stats alone, as new sections on the same single page, using the existing chart helpers, colors and card layout. Nothing comes from Moonbot, the Hub API or outside sources, so it all refreshes with the daily Job.

New sections

• Robotics: robotics datasets and models per month (models by library: LeRobot, phosphobot, openpi), 90-day and year-over-year growth, robotics' rank among model tasks per year (#40 in 2022, #2 in 2026 so far), robot platforms named in dataset tags, and the most-downloaded robotics datasets.
• Model tree: two new cards. One shows accounts publishing fine-tunes or adapters each month by base family (Qwen, Llama, Gemma…); the other is a family table with each family's mix of relations.
• Hardware formats: new models in GGUF, ONNX, MLX, NVFP4, Core ML, OpenVINO, LiteRT, ExecuTorch, MLC, AMD and NPU formats. It follows the modality chart's pattern: pick formats on a chart, plus a table with growth and share of new models.
• Size and task: shares of new models vs downloads vs likes, by parameter size and by task group.
• Evals: eval-tagged datasets per month, leaderboard and arena Spaces, and the most-liked new leaderboards. Official benchmark datasets (benchmark:official) are counted with the leaderboards.
• Science: 14 domains, models and datasets shown separately, colored by field. The domain rules are the same as evijit/hf-science-growth.

Changes to the page as a whole

• Clicking a table's column header sorts that table, including the existing tables. Values are read from the cells ("1.2K", "▼ 12%", "2.4×", months), and "Everything else" rows stay pinned at the bottom.
• The robotics, formats, science, eval and leaderboard charts have the same "Per month | Total" switch as the growth chart.

Build

• build.py gets new queries in the same DuckDB style. The only extra column read is Space tags. A local run on today's data took ~30 s (vs ~29 s before), and index.html goes from 116 KB to ~230 KB.
• Science uses build/sci_rules.py and build/data/lang_codes.json. When the Job runs build.py from its URL, it fetches both from this Space, the same way it fetches the template. DuckDB first narrows the models and datasets to candidates (~50k), and the Python rules run only on those.
• index.html and stats.json aren't in this PR; the daily Job regenerates them.