Dataset card

README.md

@@ -1,7 +1,8 @@
 ---
+pretty_name: tasksource-instruct
 language:
 - en
-license: apache-2.0
+license: other
 size_categories:
 - 1M<n<10M
 task_categories:
@@ -9,7 +10,22 @@ task_categories:
 - text-classification
 - token-classification
 - zero-shot-classification
-pretty_name: tasksource-instruct
+tags:
+- instructions
+- instruction-tuning
+- instruction-finetuning
+- flan
+- promptsource
+- tasksource
+configs:
+- config_name: default
+  data_files:
+  - split: train
+    path: data/train-*
+  - split: test
+    path: data/test-*
+  - split: validation
+    path: data/validation-*
 dataset_info:
   features:
   - name: inputs
@@ -34,65 +50,82 @@ dataset_info:
     num_examples: 193580
   download_size: 3099611926
   dataset_size: 6849144692
-tags:
-- instructions
-- instruction-tuning
-- instruction-finetuning
-- flan
-- promptsource
-- tasksource
-configs:
-- config_name: default
-  data_files:
-  - split: train
-    path: data/train-*
-  - split: test
-    path: data/test-*
-  - split: validation
-    path: data/validation-*
 ---
-# Dataset Card for "tasksource-instruct-v0" (TSI)
 
+# tasksource-instruct
 
-Multi-task instruction-tuning data recasted from 485 of the [tasksource](https://github.com/sileod/tasksource) datasets.
-Dataset size is capped at 30k examples per task to foster task diversity.
+**Instruction-tuning data recast from the ~480 English classification, multiple-choice
+and token-classification tasks of [tasksource](https://github.com/sileod/tasksource).**
+
+Every example comes from a human-built dataset (NLI, logical reasoning, sentiment,
+hate speech, discourse, argumentation, ...), not from a teacher model. Each task is
+capped at 30k training examples, so no task dominates. Many tasks aren't in FLAN v2,
+for example DynaSent, DynaHate, discriminative bAbI, epistemic logic, RuleTaker,
+veridicality and dozens of NLI datasets.
 
 ```python
-!pip install tasksource, pandit
-import tasksource, pandit
-df = tasksource.list_tasks(instruct=True).sieve(id=lambda x: 'mmlu' not in x)
-for tasks in df.id:
-  yield tasksource.load_task(task,instruct=True,max_rows=30_000,max_rows_eval=200)
+from datasets import load_dataset
+
+ds = load_dataset("tasksource/tasksource-instruct", split="train")
+ds = ds.filter(lambda use: use == "commercial", input_columns="license_use")  # optional
 ```
 
-https://github.com/sileod/tasksource
+## Format
+
+| column | content |
+|---|---|
+| `inputs` | the instruction, the example, and the answer options |
+| `targets` | the answer: an option (`entailment.`), a letter (`B.`), or `word: TAG` lines for token tasks |
+| `task` | the tasksource task id |
+| `license`, `license_use` | the source's licenses, see below |
 
-## How it differs from flan-v2
+Prompts ask for the answer with no explanation, so the short targets don't teach a
+model to stop explaining in general. Tasks are interleaved round-robin, so any
+slice of the split mixes them. Validation and test keep up to 500 examples per task.
 
-TSI is HuggingFace-centric and based on tasksource, a curated collection of HF datasets. It can be scaled to much more examples.
-tasksource is focused on discriminative tasks (Classification/TokenClassification/MultipleChoice). The coverage on discriminative tasks is greater than flan.
-List of tasks [here](https://github.com/sileod/tasksource/blob/main/tasks.md). Examples of tasks not in Flan V2 include Dynasent (adversarial sentiment analysis), Dynahate (adversarial hate speech detection, discriminative babi, epistemic logic, ruletaker, veridicality, discourse relation prediction, dozens of interesting natural language inference datasets...
+`tasksource-instruct` works well mixed with FLAN v2 or other instruction data. It
+covers discriminative reasoning tasks that those sets cover less.
 
-TSI answers are mostly short answers to multiple-choice questions, but they target a wide array of problems.
-TSI is reasoning intensive, while some flan tasks are not necessarily specific (e.g. generating hypothesis based on premise for NLI).
-We explicitly mention that answers should not have explanations, to prevent biasing models toward short answers when using other instruction datasets.
+For preference pairs built from the same rows, see
+[tasksource_dpo_pairs](https://huggingface.co/datasets/tasksource/tasksource_dpo_pairs).
+For soft labels, ratings and multi-question requests, see
+[tasksource-jev-typed-decisions](https://huggingface.co/datasets/tasksource/tasksource-jev-typed-decisions).
 
-`flan-v2` and `tasksource-instruct` can be combined to improve the reasoning capabilities of LLM. 
+## Reproducibility
 
-## Contact and citation:
-damien.sileo@inria.fr
+The dataset is built by
+[`scripts/build_instruct_dataset.py`](https://github.com/sileod/tasksource/blob/main/scripts/build_instruct_dataset.py):
 
-https://arxiv.org/abs/2301.05948
+```bash
+PYTHONPATH=.:src python scripts/build_instruct_dataset.py --finalize
 ```
+
+Sources are loaded at pinned Hub revisions. [sources.yaml](sources.yaml) records, per
+task, the Hub dataset, revision, original dataset, licenses and row counts, and
+`build-report.jsonl` records the code commit of each task's build. MMLU, BIG-bench and
+BLiMP are left out, so they stay clean for evaluation. Other public benchmarks (GLUE,
+SuperGLUE, HellaSwag, PIQA, ...) are **in** the data through their training splits.
+
+## License and scope
+
+Tasksource harmonizes datasets from many publishers; their original licenses
+and terms still apply, hence `license: other`.
+
+- `license` lists the `license` of the Hub dataset card the task was loaded from,
+  and of the original dataset behind a tasksource copy. It also lists licenses recorded
+  by the [Data Provenance Initiative](https://www.dataprovenance.org/), marked `(DPI)`.
+- `license_use` takes the most restrictive of those: `non-commercial` if any is
+  non-commercial or academic-only, `commercial` if one allows commercial use (share-alike
+  and copyleft included), and `unspecified` otherwise.
+
+This is a best-effort aid, not legal advice. Check the original terms before relying on them.
+
+## Citation
+
+```bibtex
 @inproceedings{sileo-2024-tasksource,
     title = "tasksource: A Large Collection of {NLP} tasks with a Structured Dataset Preprocessing Framework",
     author = "Sileo, Damien",
-    editor = "Calzolari, Nicoletta  and
-      Kan, Min-Yen  and
-      Hoste, Veronique  and
-      Lenci, Alessandro  and
-      Sakti, Sakriani  and
-      Xue, Nianwen",
     booktitle = "Proceedings of the 2024 Joint International Conference on Computational Linguistics, Language Resources and Evaluation (LREC-COLING 2024)",
     month = may,
     year = "2024",
@@ -101,4 +134,4 @@ https://arxiv.org/abs/2301.05948
     url = "https://aclanthology.org/2024.lrec-main.1361/",
     pages = "15655--15684",
 }
-```
+```