docs: feature as benchmark — arxiv citation, paper teaser, leaderboard cross-link

README.md

@@ -1,76 +1,129 @@
 ---
-license: cc-by-4.0
-pretty_name: SkillsBench Tasks
-tags:
-- benchmark
-- agents
-- tool-use
-- skills
+pretty_name: SkillsBench
+language:
+  - en
+license: apache-2.0
 size_categories:
-- n<1K
+  - n<1K
+task_categories:
+  - text-generation
+tags:
+  - benchmark
+  - agents
+  - tool-use
+  - skills
+  - evaluation
+  - benchflow
+arxiv: 2602.12670
+citation: |
+  @misc{skillsbench_2026,
+        title={SkillsBench: Benchmarking How Well Agent Skills Work Across Diverse Tasks},
+        author={Xiangyi Li and Wenbo Chen and Yimin Liu and Shenghan Zheng and Xiaokun Chen and Yifeng He and Yubo Li and Bingran You and Haotian Shen and Jiankai Sun and Shuyi Wang and Binxu Li and Qunhong Zeng and Di Wang and Xuandong Zhao and Yuanli Wang and Roey Ben Chaim and Zonglin Di and Yipeng Gao and Junwei He and Yizhuo He and Liqiang Jing and Luyang Kong and Xin Lan and Jiachen Li and Songlin Li and Yijiang Li and Yueqian Lin and Xinyi Liu and Xuanqing Liu and Haoran Lyu and Ze Ma and Bowei Wang and Runhui Wang and Tianyu Wang and Wengao Ye and Yue Zhang and Hanwen Xing and Yiqi Xue and Steven Dillmann and Han-chung Lee},
+        year={2026},
+        eprint={2602.12670},
+        archivePrefix={arXiv},
+        primaryClass={cs.AI},
+        url={https://arxiv.org/abs/2602.12670}}
 configs:
-- config_name: default
-  data_files:
-  - split: train
-    path: skillsbench-tasks.parquet
+  - config_name: default
+    data_files:
+      - split: train
+        path: skillsbench-tasks.parquet
 ---
 
-# SkillsBench Tasks
+> **Read-only mirror.** The primary source for these tasks is GitHub at
+> [`benchflow-ai/skillsbench`](https://github.com/benchflow-ai/skillsbench). Open issues
+> and pull requests there.
+>
+> **Leaderboard / trajectories:** [`benchflow/skillsbench-leaderboard`](https://huggingface.co/datasets/benchflow/skillsbench-leaderboard).
+
+# SkillsBench
+
+> Agent Skills are structured packages of procedural knowledge that augment LLM agents
+> at inference time. SkillsBench measures how well agents put those skills to work on
+> realistic professional tasks across many domains.
+
+- **Paper:** [arXiv:2602.12670](https://arxiv.org/abs/2602.12670)
+- **Code & tasks:** [`github.com/benchflow-ai/skillsbench`](https://github.com/benchflow-ai/skillsbench)
+- **Runner (BenchFlow SDK):** [`github.com/benchflow-ai/benchflow`](https://github.com/benchflow-ai/benchflow)
+- **Discord:** [`discord.gg/G9dg3EfSva`](https://discord.gg/G9dg3EfSva)
+
+## Quickstart
+
+The canonical install + run instructions live in the github repo's [README](https://github.com/benchflow-ai/skillsbench#readme)
+and [CONTRIBUTING.md](https://github.com/benchflow-ai/skillsbench/blob/main/CONTRIBUTING.md).
+At the time of this snapshot the runner is invoked as `bench eval create`; check the github
+repo for the current install command before copying any line below.
+
+```bash
+# Run the oracle agent (no API key required) to confirm tasks load:
+bench eval create -t tasks --agent oracle
 
-Full task corpus from [`benchflow-ai/skillsbench`](https://github.com/benchflow-ai/skillsbench)
-bundled into a single parquet for offline analysis.
+# Run a real agent with skills:
+bench eval create \
+  -t tasks --agent claude-agent-acp -m claude-opus-4-7 \
+  -s tasks/<task-id>/environment/skills \
+  -o jobs/with-skills
 
-**Source commit:** `05f263429c5033604da64e8de568ce654fcbe6e6`
-**Task count:** 91
-**Scope:** full corpus including `solution/` and `tests/`. **Do not** train models on
-this dataset — solutions and ground-truth tests are included for reproducibility, and
-contamination on those will void any benchmark numbers reported against SkillsBench.
+# … and again without skills, to measure the skill delta:
+bench eval create \
+  -t tasks --agent claude-agent-acp -m claude-opus-4-7 \
+  -o jobs/no-skills
+```
+
+To submit a run to the leaderboard, copy `jobs/<run-timestamp>/` into a fork of
+[`benchflow/skillsbench-leaderboard`](https://huggingface.co/datasets/benchflow/skillsbench-leaderboard)
+under `submissions/skillsbench/v0.1/<agent>__<model>/` and open a PR. See that repo's
+README for the full submission protocol.
 
-## Schema
+## What's in this dataset
 
-One row per task. Top-level columns:
+A single parquet bundling the full task corpus from upstream `main` at commit
+`05f263429c5033604da64e8de568ce654fcbe6e6` (91 tasks). One row per task with
+`task.toml` metadata flattened to top-level columns, key text files inlined
+(`instruction`, `dockerfile`, `solve_sh`, `test_sh`, `test_outputs`), and every
+other file (environment data, solution artefacts, test fixtures, skill assets) in
+`files: list[struct{path, size_bytes, sha256, is_text, content}]` with binary
+content as parquet `binary`.
 
-| column | type | description |
+| Field | Type | Description |
 |---|---|---|
 | `task_id` | string | directory name under `tasks/` |
-| `category` | string | task.toml `[metadata].category` |
-| `difficulty` | string | `easy` / `medium` / `hard` |
+| `category`, `difficulty` | string | from `task.toml` `[metadata]` |
 | `difficulty_explanation` | string | author's prose on why the task is hard |
-| `author_name`, `author_email` | string | task author |
 | `tags` | list[string] | `[metadata].tags` |
 | `allow_internet` | bool | `[environment].allow_internet` |
-| `verifier_timeout_sec`, `agent_timeout_sec`, `build_timeout_sec` | float | task.toml timeouts |
-| `cpus`, `memory_mb`, `storage_mb` | int | environment limits |
-| `instruction` | string | full `instruction.md` text |
-| `task_toml` | string | raw `task.toml` |
-| `dockerfile` | string | `environment/Dockerfile` |
-| `solve_sh` | string | `solution/solve.sh` |
-| `test_sh`, `test_outputs` | string | `tests/test.sh`, `tests/test_outputs.py` |
+| `instruction`, `task_toml`, `dockerfile`, `solve_sh`, `test_sh`, `test_outputs` | string | inlined source files |
 | `skills` | list[struct] | `{name, description, skill_md}` per skill in `environment/skills/` |
-| `files` | list[struct] | every other file in the task tree: `{path, size_bytes, sha256, is_text, content}` |
-
-`files` covers `environment/`, `solution/`, `tests/`, and skill assets (excluding the SKILL.md
-already captured in `skills`). Binary blobs (xlsx, mp4, zip, mseed, parquet) are stored in
-`content` as bytes; `is_text` indicates whether `content.decode("utf-8")` is meaningful.
-
-## Reading
+| `files` | list[struct] | every other file: `{path, size_bytes, sha256, is_text, content}` |
 
 ```python
 from huggingface_hub import hf_hub_download
-import pyarrow.parquet as pq
+import pyarrow.parquet as pq, pyarrow.compute as pc
 
 path = hf_hub_download("benchflow/skillsbench",
                        "skillsbench-tasks.parquet", repo_type="dataset")
 t = pq.read_table(path)
-print(t.num_rows, t.column_names)
-
-# Reconstitute one task on disk:
-row = t.filter(pa.compute.equal(t["task_id"], "shock-analysis-demand")).to_pylist()[0]
+row = t.filter(pc.equal(t["task_id"], "shock-analysis-demand")).to_pylist()[0]
 for f in row["files"]:
     ...  # write f["content"] to f["path"]
 ```
 
-## License
+## Anti-contamination
+
+> **Do not train models on this dataset.** Solutions and ground-truth tests are inlined
+> for reproducibility. Contamination on those will void any benchmark numbers reported
+> against SkillsBench.
 
-Source code under Apache-2.0 (per upstream `LICENSE`); the dataset itself is published
-under CC-BY-4.0 to match the NeurIPS Datasets & Benchmarks track standard.
+## Citing SkillsBench
+
+```bibtex
+@misc{skillsbench_2026,
+      title={SkillsBench: Benchmarking How Well Agent Skills Work Across Diverse Tasks},
+      author={Xiangyi Li and Wenbo Chen and Yimin Liu and Shenghan Zheng and Xiaokun Chen and Yifeng He and Yubo Li and Bingran You and Haotian Shen and Jiankai Sun and Shuyi Wang and Binxu Li and Qunhong Zeng and Di Wang and Xuandong Zhao and Yuanli Wang and Roey Ben Chaim and Zonglin Di and Yipeng Gao and Junwei He and Yizhuo He and Liqiang Jing and Luyang Kong and Xin Lan and Jiachen Li and Songlin Li and Yijiang Li and Yueqian Lin and Xinyi Liu and Xuanqing Liu and Haoran Lyu and Ze Ma and Bowei Wang and Runhui Wang and Tianyu Wang and Wengao Ye and Yue Zhang and Hanwen Xing and Yiqi Xue and Steven Dillmann and Han-chung Lee},
+      year={2026},
+      eprint={2602.12670},
+      archivePrefix={arXiv},
+      primaryClass={cs.AI},
+      url={https://arxiv.org/abs/2602.12670}}
+```