Add files using upload-large-folder tool

README.md

@@ -113,6 +113,57 @@ dataset, task_id, category, task_file, asset_dir, prompt_files, skill_count
 
 The full task definitions remain in the YAML files.
 
+## Responsible AI (RAI) Considerations
+
+This section summarizes the Responsible AI information also encoded in the Croissant metadata under
+`croissant/openreview_croissant.json`.
+
+### Synthetic Data Generation
+
+ClawBenchPro is designed as a synthetic workplace-agent benchmark. Tasks were generated, staged,
+imported into Nanoclaw task YAML format, validated with task-local `env_builder.py` builders, grouped
+into aligned benchmark subsets, and exported in compact builder-only form for hosting. The benchmark
+includes fictional workplace scenarios, synthetic personas, synthetic internal services, synthetic
+records, task-local skills, and generated workspace fixtures. The package is not intended to contain
+real user records or operational credentials.
+
+### Sandbox Fixtures and Real-World Safety
+
+Some tasks intentionally contain strings that look like API keys, internal URLs, secrets, policy-sensitive
+phrases, corrupted files, or broken logs. These are sandbox benchmark fixtures used to test whether agents
+can reason about noisy workplace environments. They are not real credentials, do not connect to production
+systems, and are meant to be executed in local or containerized workspaces. Prebuilt `assets/` are not
+published; environments are materialized locally from `env_builder.py` when needed.
+
+### Environment Validation Limitations
+
+Environment builders were checked for the ability to materialize task workspaces, but this does not prove
+that every possible runner, model, operating system, dependency version, or execution policy will behave
+identically. The benchmark measures performance in controlled Nanoclaw-compatible environments and should
+not be interpreted as a complete certification of real-world workplace automation reliability.
+
+### Biases and Scope Limitations
+
+The dataset is selected to stress agentic workflows such as state tracking, tool use, multi-turn reasoning,
+file manipulation, and skill invocation. This creates selection bias toward tasks that can be packaged as
+local workspaces. The benchmark may underrepresent low-resource languages, non-technical occupations,
+accessibility-specific workflows, embodied tasks, and domains requiring direct human interaction.
+
+### Intended and Non-Recommended Uses
+
+Intended uses include comparative evaluation of language-agent runners, category-level benchmark analysis,
+reproducibility studies, and diagnosis of tool-use or multi-turn failure modes. The dataset is not validated
+for model fine-tuning, safety certification, demographic fairness auditing, hiring decisions, medical,
+legal, financial, or other safety-critical deployment decisions.
+
+### Social Impact
+
+ClawBenchPro may help make agent evaluation more reproducible and transparent by providing local synthetic
+workspaces, explicit task categories, and machine-readable metadata. Potential risks include overgeneralizing
+benchmark scores, optimizing narrowly to synthetic tasks, or presenting benchmark performance as evidence of
+real-world reliability. We mitigate these risks by documenting limitations, keeping tasks sandboxed, and
+publishing Croissant metadata with RAI fields.
+
 ## License
 
 This package is released under the MIT License. See `LICENSE`.

checksums.sha256

@@ -2,7 +2,7 @@
 5b56373d7af94293eff83e888fdbfe1f40ebdb41d2724c3475671e10e3d86070  .gitignore
 5b56373d7af94293eff83e888fdbfe1f40ebdb41d2724c3475671e10e3d86070  .hfignore
 4e7ce122f89dff36d79ce68019940fb84c4824862e07381955b603364a421beb  LICENSE
-f608aa6e2b74f8762b17818d5af6c8a6322ed30932ad88704e41f32e29088485  README.md
+759516519b1c2180469b3ce9ec8d16b15242126c5999d989c9d3c9a4483393d1  README.md
 7f634c0f727e8ce82e241586d3524fbc5e6ec1d8dccf8bd81c7db897739a8137  croissant/clawbenchpro_collection_croissant.json
 470c0df27cc256f3be7927090df53bd395b1a5c2b95bb992930fd1b5022ba1d3  croissant/huggingface_auto_croissant.json
 0eaef57248c7e718f56188c66d82128d4c05f7912bad8b9e6b125c3adb775ecd  croissant/openreview_croissant.json
@@ -10,7 +10,7 @@ a39dfb9069001a2190d69bde656d6a5695ee6e3f37aadb95c20c549ecfa33f18  croissant/pers
 6ae589ec7e7f6edc25d03abe224a65143be887aa520cd98d687b8269758366e5  croissant/round_01_aligned_mix_800_croissant.json
 aa1265325f58610ba9864647d132772414a9c6b594dc5e9b688c2a6d316b0f43  dataset_index.csv
 c40acaaf392927804d85b7496afeb13135205dfbd8828c562b1c6866431ff151  dataset_index.jsonl
-b1c929f4988e9ec2e172a2b0779cb8d15a366363e57802823086fdc90a633015  manifest.json
+5abe432105c8a8960396d5242c61c84ea956ad81468bc8297b71311937ed3762  manifest.json
 011d2f2a47ada6ad72d66bd8fd275ea7a8ee6ef8301fdede1f6b9eef2400297d  materialize_assets.py
 1d2f56b5d90ce106996e79ff7563c08d4855bcc67924c2ebc67b77b6eb50d14e  persona_aligned_mix_200/README.md
 bc557847f6a9703665bee383921cde25d8ee4a0c771286686935aa5c7fa0b9bb  persona_aligned_mix_200/checksums.sha256

manifest.json

@@ -1,6 +1,6 @@
 {
   "package_name": "ClawBenchPro",
-  "created_at": "2026-05-06T09:42:37.188718+00:00",
+  "created_at": "2026-05-06T09:50:45.548384+00:00",
   "source_package": "published_datasets/nanoclaw_workplace_suite",
   "format": "huggingface_dataset_repository",
   "license": "MIT",