LQ-Decide training data: 141k typed decisions, licence-tagged sources, synthetic families

.gitattributes

@@ -58,3 +58,5 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 # Video files - compressed
 *.mp4 filter=lfs diff=lfs merge=lfs -text
 *.webm filter=lfs diff=lfs merge=lfs -text
+families.jsonl filter=lfs diff=lfs merge=lfs -text
+train.jsonl filter=lfs diff=lfs merge=lfs -text

README.md

@@ -0,0 +1,106 @@
+---
+license: cc-by-4.0
+task_categories:
+  - text-classification
+  - zero-shot-classification
+language:
+  - en
+size_categories:
+  - 100K<n<1M
+tags:
+  - typed-decisions
+  - decision-model
+  - calibration
+  - routing
+---
+
+# LQ-Decide training data
+
+141,038 rows for training models that answer **typed decisions**: given a state and a question with a fixed option set,
+return a probability over the options rather than generated text. Built for
+[LQ-Decide 0.6B](https://huggingface.co/Hanish/lq-decide-0.6b) by [Hanish Keloth](https://huggingface.co/Hanish).
+
+Every source is licence-checked and named. Non-commercial and unclear-licence sources were excluded by a flag rather
+than being quietly included; the excluded list is below so you can decide for yourself.
+
+## Files
+
+| file | rows | contents |
+|---|---|---|
+| `train.jsonl` | 109,523 | public datasets recast as typed decisions |
+| `valid.jsonl` | 2,235 | held-out split of the same mixture |
+| `families.jsonl` | 29,280 | synthetic rows in three decision families |
+
+## Row format
+
+```json
+{
+  "id": "65d834b2683fd321",
+  "family": "evidence_interpretation",
+  "kind": "choice",
+  "state": "The optician ordered replacement lenses. The workshop confirms they have not yet been fitted.",
+  "question": "Assess the claim: the replacement lenses have been fitted.",
+  "options": [
+    {"id": "supported",    "description": "The evidence establishes the claim"},
+    {"id": "insufficient", "description": "The evidence does not establish either"},
+    {"id": "contradicted", "description": "The evidence establishes the opposite"}
+  ],
+  "label": 2,
+  "source": "authored_synthetic",
+  "licence": "generated by Ornith-1.5-35B-A3B (MIT)"
+}
+```
+
+`kind` is `choice`, `noul` (yes/no) or `score` (ordered levels). `label` indexes `options`. Option order is already
+shuffled per row, so a model trained on this reads descriptions rather than positions.
+
+## Sources in `train.jsonl`
+
+| source | rows | licence |
+|---|---|---|
+| MNLI | 20,000 | OANC + CC-BY-SA (mixed) |
+| SNLI | 20,000 | CC-BY-SA 4.0 |
+| WANLI | 20,000 | CC-BY 4.0 |
+| CLINC150 | 15,250 | CC-BY 3.0 |
+| MASSIVE (mteb mirror) | 11,514 | CC-BY 4.0 |
+| Banking77 (mteb mirror) | 9,993 | CC-BY 4.0 |
+| BoolQ | 9,427 | CC-BY-SA 3.0 |
+| SMS Spam (UCI) | 5,574 | CC-BY 4.0 |
+
+Deliberately excluded: ANLI (CC-BY-NC), DAIR Emotion (research only), AG News (unclear), Amazon Reviews 2023 (custom
+research licence), papluca language-identification (derived, unclear). Regenerate with or without them using the script
+referenced below.
+
+## `families.jsonl`
+
+29,280 synthetic rows across three families, fifteen mechanisms each, balanced across answers:
+
+| family | rows | answers |
+|---|---|---|
+| evidence interpretation | 9,880 | supported / insufficient / contradicted |
+| candidate selection | 9,825 | Candidate A / Candidate B / neither |
+| rule application | 9,575 | permitted / prohibited / unsettled |
+
+Mechanisms include authority hierarchy, delegation limits, exception and exclusion, partial deliverables, priority
+fallback, revocation, timing sequence, and family-specific ones such as approval stage and retained-versus-replaced.
+
+**Generation method.** The family, mechanism, intended answer and domain were chosen first, and
+`Ornith-1.5-35B-A3B` (MIT) wrote only the surface text for that predetermined answer. The label therefore never came
+from the teacher, which removes teacher-labelling noise. Rows whose returned answer did not match the requested one were
+discarded; yield was 98.6%. Sixty professional domains are used to spread vocabulary.
+
+**Known weakness, measured.** A model trained on this reaches 0.646 on candidate selection on a clean held-out fixture
+but 0.389 on perturbed variants of the same items, which says these generated candidate-selection rows are too
+formulaic and teach a surface cue. The other two families hold up much better. If you use this data, consider
+regenerating that family with harder negatives.
+
+## Evaluation
+
+Not included here. Use the labelled fixtures from the [SemIf](https://github.com/TheoLeeCJ/openjev) project
+(`authored144`, `perturbations108`, MIT). They were never trained on, generated from, or shown to the teacher when this
+data was built.
+
+## Reproduce
+
+Built by `ml/decide/build_data.py` (public sources) and `ml/modal/gen_decide_families.py` (synthetic families) in the
+LocalQuill repository. The first takes a `--commercial-only` flag that produces exactly the source list above.

families.jsonl

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4856411d56c76c17a3339d41fe55cd869fdbe7d3e156ec67beb5f0fbb10e9b48
+size 29368917

train.jsonl

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:02b0dfc0c033095ced72d382dbeb892d7787fba48943a2e93ee36524432d9431
+size 88198869