Add UnCoBench dataset card

Document task settings, data structure, usage, evaluation and limitations; align the data license with CC BY 4.0.

README.md

@@ -1,3 +1,127 @@
 ---
-license: cc-by-nc-4.0
+license: cc-by-4.0
+language:
+- en
+pretty_name: UnCoBench
+size_categories:
+- 1K<n<10K
+task_categories:
+- text-generation
+tags:
+- text
+- benchmark
+- multi-agent
+- collaboration
+- reasoning
+- synthetic
+- evaluation
 ---
+
+# UnCoBench
+
+[GitHub Repository](https://github.com/iNLP-Lab/UnCoBench)
+
+UnCoBench evaluates how well a main agent completes collaborative tasks when other agents exhibit unreliable behavior. The dataset contains **1,000 groups of paired English tasks** across **five settings**, yielding **5,000 task instances**.
+
+The main agent must gather information from role agents, resolve gaps or inconsistencies, and produce a final decision and, when required, a numerical result. Role agents have separate descriptions and private information. Shared task IDs make it possible to compare performance across collaboration settings.
+
+## Dataset Overview
+
+| Property | Description |
+| --- | --- |
+| Language | English |
+| Task groups | 1,000, identified by `task_001` through `task_1000` |
+| Task instances | 5,000: 1,000 in each setting |
+| Interaction | A main agent queries role agents and submits a final answer |
+| Release format | Task folders containing a Python runtime and role-agent JSON files |
+| Evaluation | Reference-based correctness and communication metrics computed by the task runtime |
+
+The five settings are evaluation conditions, not training, validation, or test splits. The release does not define a separate train/dev/test partition.
+
+## Collaboration Settings
+
+The four primary settings support paired comparisons of individual unreliable behaviors. The additional `mixed` setting combines them as a stress test.
+
+| Setting | Description |
+| --- | --- |
+| `reliable` | Baseline collaboration without the three injected unreliable behaviors. |
+| `misaligned` | A required record in the public candidate directory points to a different context. The appropriate record remains available, requiring the main agent to recognize and resolve the mismatch. |
+| `incomplete` | A necessary numerical answer is initially omitted or given as a truthful but insufficient bound or interval. A follow-up can recover the precise value. |
+| `missed` | A required role initially does not respond. The main agent must follow up in a later round to obtain an answer. |
+| `mixed` | Combines the three unreliable behaviors across three different required roles. |
+
+`reliable` denotes the absence of injected interventions; it does not guarantee that the models will answer correctly.
+
+## Data Structure
+
+```text
+data/
+  reliable/
+    task_001/
+      tasks/
+        run_one.py
+      role_agents/
+        facility_utilities.json
+        ...
+    ...
+    task_1000/
+  misaligned/
+  incomplete/
+  missed/
+  mixed/
+```
+
+Each setting follows the same task-folder layout:
+
+- **`tasks/run_one.py`** contains the task query, reference answer, interaction protocol, setting-specific behavior, and evaluation logic.
+- **`role_agents/*.json`** contains each role's description and private information. Fields include `role_agent_name`, `abstract`, `system_message`, and `information_points`; each information point has a `title` and `content`.
+
+Reference answers and private role information are available to the evaluator. The main agent should obtain role information through the interaction protocol rather than reading these files directly.
+
+## Download
+
+Use `huggingface_hub` to download the repository while preserving the task-folder structure. Authenticate with `hf auth login` if required by the repository's access settings.
+
+```bash
+python -m pip install -U huggingface_hub
+```
+
+```python
+from huggingface_hub import snapshot_download
+
+snapshot_download(
+    repo_id="iNLP-Lab/UnCoBench",
+    repo_type="dataset",
+    local_dir="UnCoBench",
+)
+```
+
+This release contains executable task folders and role definitions rather than a single table of independent question-answer rows.
+
+## Evaluation
+
+The [GitHub repository](https://github.com/iNLP-Lab/UnCoBench) provides the batch evaluation runner, dependencies, and setup instructions. It supports both [API-based main models](https://github.com/iNLP-Lab/UnCoBench#closed-source-model-evaluation) and [locally hosted main models](https://github.com/iNLP-Lab/UnCoBench#open-sourced-model-evaluation), with role agents configured separately. API credentials are supplied through environment variables or a local `.env` file.
+
+During evaluation, the main agent exchanges JSON messages with role agents and submits `conclusion` and `numeric_result`. Correctness requires the conclusion and the applicable numerical result to match the task's reference answer, using the task's numerical precision rules. The runtime also records communication measures such as turns to finalization and recovery of missing or misaligned information. Scoring is computed from reference answers and the interaction transcript; it does not use an additional LLM judge.
+
+For comparisons, keep the role model and evaluation configuration fixed, and compare matching task IDs across settings. Report the main and role model versions, selected settings, task range, and interaction-budget settings alongside results.
+
+## Dataset Construction
+
+UnCoBench consists of synthetic collaborative tasks. The accompanying [generation pipeline](https://github.com/iNLP-Lab/UnCoBench/tree/main/pipeline) uses language models to construct scenario semantics and natural-language descriptions, while code compiles task constraints, role information, and record routing. A deterministic solver derives the reference answers.
+
+The pipeline includes a full-information qualification step and derives paired variants for the five collaboration settings. The data release contains the resulting task runtimes and role definitions; the generation code is maintained in the GitHub repository.
+
+## Intended Use and Limitations
+
+UnCoBench is intended for research on information gathering, follow-up behavior, contextual consistency, and decision-making in multi-agent systems.
+
+- Tasks are synthetic and use a defined interaction protocol. Results should not be treated as a direct estimate of performance in unrestricted real-world collaboration.
+- Outcomes depend on both the main model and the models used for role agents, as well as their generation parameters and turn budgets.
+- Private role information and reference answers must remain outside the main agent's direct inputs to preserve the intended information asymmetry.
+
+## License and Attribution
+
+Benchmark data, including task content embedded in `run_one.py`, is licensed under [CC BY 4.0](https://creativecommons.org/licenses/by/4.0/), consistent with the repository's [data license](https://github.com/iNLP-Lab/UnCoBench/blob/main/LICENSE-DATA). Code, documentation, and generation prompts are covered by the repository's [MIT license](https://github.com/iNLP-Lab/UnCoBench/blob/main/LICENSE).
+
+When using or redistributing the data, credit the UnCoBench contributors, link to the [dataset](https://huggingface.co/datasets/iNLP-Lab/UnCoBench) and [code repository](https://github.com/iNLP-Lab/UnCoBench), include the license, and indicate any modifications.