Datasets:
File size: 5,309 Bytes
36db5d6 9867c75 36db5d6 9867c75 36db5d6 9867c75 36db5d6 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 | ---
license: cc-by-nc-4.0
language:
- en
task_categories:
- text-retrieval
pretty_name: MoraLink
size_categories:
- n<1K
tags:
- retrieval
- benchmark
- fables
- morals
- narrative
configs:
- config_name: fables
data_files:
- split: test
path: data/fables.json
- config_name: queries
data_files:
- split: test
path: data/queries.json
- config_name: qrels
data_files:
- split: test
path: data/qrels.json
- config_name: clusters
data_files:
- split: test
path: data/clusters.json
- config_name: summaries-gemini-3-flash
data_files:
- split: test
path: summaries/gemini-3-flash.json
- config_name: summaries-gemma-4-e2b
data_files:
- split: test
path: summaries/gemma-4-e2b.json
- config_name: summaries-gemma-4-e4b
data_files:
- split: test
path: summaries/gemma-4-e4b.json
- config_name: summaries-gemma-4-31b
data_files:
- split: test
path: summaries/gemma-4-31b.json
---
# MoraLink
MoraLink is a moral-to-fable retrieval benchmark. Given a short moral lesson, a retrieval system must rank the fables that express that lesson.
- Paper: [MoraLink: Bridging Morals and Narrative Fables for Retrieval](https://openreview.net/forum?id=wpUUxOAzZY), EMNLP 2026
- Code and reproduction instructions: [Intellexus-DSI/MoraLink](https://github.com/Intellexus-DSI/MoraLink)
MoraLink is derived from the English MORABLES dataset and contains:
- 709 fables
- 668 unique moral queries
- 558 moral groups
- 1,085 query-to-fable relevance labels
Some moral queries have one relevant fable and others have several. The benchmark therefore uses multi-positive relevance labels.
## Main files
### Benchmark data
- `data/fables.json`: The 709 fable documents. Each row contains `doc_id`, `title`, `text`, and the original MORABLES alias.
- `data/queries.json`: The 668 moral queries. Each row contains the query text, moral-group information, and relevant fable IDs.
- `data/qrels.json`: The 1,085 query-to-fable relevance labels used for evaluation.
- `data/clusters.json`: The final 558 moral groups.
- `data/moral_to_cluster.json`: The mapping from original moral records to the final moral groups.
### Generated summaries
The `summaries/` directory contains the four frozen 709-fable summary collections used in the MoraLink experiments:
- Gemini 3 Flash
- Gemma 4 E2B
- Gemma 4 E4B
- Gemma 4 31B
Each summary file contains the prompt variants generated for every fable. The exact prompts, model identifiers, generation settings, file hashes, and experiment mapping are in `metadata/summary-prompts.json` and `metadata/summary-manifest.json`.
The Gemini file preserves one API error placeholder for `gibbs_137_584`. This is intentional because the reported experiments used that frozen file. See the manifest for details.
### Annotation records
The `annotation/` directory contains the moral-grouping protocol, the surviving accepted and rejected decisions, the 23 documented moral-text edits, and the original-annotator confirmations used to finalize the release.
### Rankings and analysis files
- `rankings/manifest.json`: public metadata for the 144 official zero-shot runs.
- `rankings/*.json`: compact top-10 rankings for all 144 runs.
- `analysis/per_query_paper_metrics.csv.gz`: 96,192 per-query metric rows used to reproduce the paper-compatible MRR and the one-versus-several analysis.
- `analysis/word_overlap_rank_changes.csv.gz`: 106,212 per-query rows used to reproduce the 159 query-summary word-overlap comparisons.
The compact rankings do not include every document below rank 10. The per-query paper-metric file stores the reciprocal-rank contribution needed when the first relevant fable is below rank 10.
## Loading the data
```python
from datasets import load_dataset
fables = load_dataset("Intellexus/MoraLink", "fables", split="test")
queries = load_dataset("Intellexus/MoraLink", "queries", split="test")
qrels = load_dataset("Intellexus/MoraLink", "qrels", split="test")
summaries = load_dataset(
"Intellexus/MoraLink",
"summaries-gemini-3-flash",
split="test",
)
```
The generated-summary files use the original MORABLES alias in `original_fable_id`. Join this field with the `alias` field in the fable collection.
## Reproducibility
All released files are frozen copies of the files used in the reported experiments. `FILE_HASHES.sha256` contains a SHA-256 hash for every released file.
The release was validated as four summary collections with 709 rows each, 144 compact rankings with 668 queries each, and 558 moral groups, including 469 singleton groups and 89 groups with several fables. There are no unresolved annotation records.
## License and source
MoraLink is derived from [MORABLES](https://huggingface.co/datasets/cardiffnlp/Morables) and follows its CC BY-NC 4.0 license. See `LICENSE.md` and `NOTICE.md` for the full reuse conditions, source attribution, and generator information.
Generated summaries may also be subject to the terms of the named model or service provider.
## Scope
MoraLink studies English-language Western fable traditions represented in MORABLES. It should not be treated as a general benchmark of moral reasoning across languages or cultures.
|