README.md

---
pretty_name: ClosureBench
language:
- en
license: cc-by-4.0
task_categories:
- text-classification
- question-answering
tags:
- benchmark
- llm-evaluation
- reasoning
- knowledge-representation
- open-world-assumption
- closed-world-assumption
- local-closed-world-assumption
- agent-evaluation
- symbolic-reasoning
configs:
- config_name: base
  data_files:
  - split: full
    path: data/base/full.jsonl
  - split: train
    path: data/base/train.jsonl
  - split: validation
    path: data/base/validation.jsonl
  - split: test
    path: data/base/test.jsonl
- config_name: ask_act
  data_files:
  - split: full
    path: data/ask_act/full.jsonl
  - split: train
    path: data/ask_act/train.jsonl
  - split: validation
    path: data/ask_act/validation.jsonl
  - split: test
    path: data/ask_act/test.jsonl
- config_name: multi_agent
  data_files:
  - split: full
    path: data/multi_agent/full.jsonl
  - split: train
    path: data/multi_agent/train.jsonl
  - split: validation
    path: data/multi_agent/validation.jsonl
  - split: test
    path: data/multi_agent/test.jsonl
- config_name: dynamic_dialogue
  data_files:
  - split: full
    path: data/dynamic_dialogue/full.jsonl
  - split: train
    path: data/dynamic_dialogue/train.jsonl
  - split: validation
    path: data/dynamic_dialogue/validation.jsonl
  - split: test
    path: data/dynamic_dialogue/test.jsonl
---

# ClosureBench

ClosureBench is a controlled benchmark for evaluating if LLMs
respect explicit semantic contracts about missing information. It tests whether
models distinguish absence-as-unknown, absence-as-false, and
absence-as-false-only-in-complete-scopes under explicit open-world,
closed-world, and locally closed-world contracts.

The dataset includes the base benchmark and three extensions:

| Config | `full` rows | Description |
|---|---:|---|
| `base` | 960 | Main OWA/CWA/LCWA benchmark with fixed facts, rules, and query across semantic variants. |
| `ask_act` | 960 | Maps truth values to operational actions: `approve`, `deny`, or `request_information`. |
| `multi_agent` | 360 | Tests whether a coordinator preserves source-scoped closure. |
| `dynamic_dialogue` | 100 | Tests whether models update conclusions when complete predicates change across turns. |

Each config exposes four splits:

| Split | Meaning |
|---|---|
| `full` | Full split used for the paper's reported metrics. |
| `train` | Internal benchmark train partition. |
| `validation` | Internal benchmark development partition. |
| `test` | Internal benchmark test partition. |

## Loading

```python
from datasets import load_dataset

base = load_dataset("ML0037/ClosureBench", "base", split="full")
ask_act = load_dataset("ML0037/ClosureBench", "ask_act", split="full")
multi_agent = load_dataset("ML0037/ClosureBench", "multi_agent", split="full")
dynamic = load_dataset("ML0037/ClosureBench", "dynamic_dialogue", split="full")
```

For the held-out partition only:

```python
base_test = load_dataset("ML0037/ClosureBench", "base", split="test")
```

## Base Results

Values are three-run mean +/- sample standard deviation on the `base` config,
`full` split.

| Model | Semantic switch | Core switch | LCWA closed | LCWA open | Overall |
|---|---:|---:|---:|---:|---:|
| Mistral Small | 55.73 +/- 2.43 | 27.60 +/- 3.65 | 31.60 +/- 4.34 | 31.25 +/- 1.80 | 80.38 +/- 1.02 |
| DeepSeek Flash | 81.25 +/- 0.83 | 68.75 +/- 1.38 | 49.65 +/- 3.18 | 100.00 +/- 0.00 | 90.31 +/- 0.47 |
| DeepSeek Pro | 86.46 +/- 0.18 | 77.43 +/- 0.30 | 62.15 +/- 1.20 | 100.00 +/- 0.00 | 93.61 +/- 0.21 |
| Mistral Medium | 87.08 +/- 0.65 | 78.47 +/- 1.09 | 56.95 +/- 2.17 | 100.00 +/- 0.00 | 95.59 +/- 0.37 |
| Llama Scout | 50.73 +/- 0.48 | 18.23 +/- 1.04 | 4.51 +/- 1.59 | 52.43 +/- 3.01 | 71.18 +/- 0.53 |
| Llama Maverick | 99.59 +/- 0.18 | 99.31 +/- 0.30 | 99.65 +/- 0.60 | 100.00 +/- 0.00 | 99.86 +/- 0.06 |

Semantic switch accuracy is the primary metric: a base scenario is correct only
when all semantic variants of that scenario are answered correctly.

## Result Artifacts

The `results/scored/` directory contains final `*_scored.jsonl` files used to
compute reported metrics. Raw provider response dumps are intentionally not
included.

The `results/reports/` directory contains JSON manifests and aggregate
summaries, including:

| File | Purpose |
|---|---|
| `results/reports/closurebench_replicate_summary.json` | Base benchmark three-run summary. |
| `results/reports/closurebench_replicate_manifest.json` | Base benchmark scored-run manifest. |
| `results/reports/closurebench_model_comparison.json` | Single-run base comparison. |
| `results/reports/closurebench_ask_act_summary.json` | Ask/Act extension summary. |
| `results/reports/closurebench_multi_agent_summary.json` | Multi-Agent extension summary. |
| `results/reports/closurebench_dynamic_dialogue_summary.json` | Dynamic Dialogue extension summary. |

## Data Fields

Common fields include:

- `id`: item identifier.
- `base_id`: contrastive scenario identifier.
- `split`: original benchmark partition (`train`, `dev`, or `test`).
- `domain`, `family`, `subset`: item grouping metadata.
- `semantics`: semantic contract for base-style items (`owa`, `cwa`, `lcwa`).
- `closed_predicates`: predicates declared complete for the item.
- `facts_positive`, `facts_negative`, `rules_natural`: natural-language KB.
- `symbolic`: symbolic atoms, rules, query atom, and closure atoms.
- `prompt`: exact prompt used for evaluation.
- `gold_answer`, `gold_truth_value`, or extension-specific gold fields.

Extension configs add task-specific fields, such as `gold_action` for Ask/Act,
`gold_source_used` for Multi-Agent, and turn-level gold labels for Dynamic
Dialogue.

## Limitations

ClosureBench is a targeted diagnostic benchmark. It isolates closure-contract
compliance under controlled prompts but it is not a broad measure of general agent
performance, factual knowledge, or end-to-end tool-use reliability.


## License

The dataset is released under CC BY 4.0.