Publish verified augmented SFT train, dev, and test splits

.gitattributes

@@ -58,3 +58,6 @@ 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
+dev.jsonl filter=lfs diff=lfs merge=lfs -text
+test.jsonl filter=lfs diff=lfs merge=lfs -text
+train.jsonl filter=lfs diff=lfs merge=lfs -text

README.md

@@ -1,3 +1,115 @@
 ---
-license: apache-2.0
+pretty_name: EGSciQA-ptPT-V1
+language:
+- pt
+task_categories:
+- text-generation
+- question-answering
+tags:
+- scientific-reasoning
+- grounded-generation
+- supervised-fine-tuning
+- portuguese
+size_categories:
+- 10K<n<100K
+configs:
+- config_name: default
+  data_files:
+  - split: train
+    path: train.jsonl
+  - split: validation
+    path: dev.jsonl
+  - split: test
+    path: test.jsonl
 ---
+
+# EGSciQA-ptPT-V1
+
+EGSciQA-ptPT-V1 is a European Portuguese (`pt-PT`) supervised fine-tuning
+dataset for evidence-grounded scientific question answering and reasoning. Each
+example contains a system prompt, an instruction with scientific evidence, and
+a structured target response using `<raciocinio>` and `<resposta>` sections.
+
+This repository contains the final instruction-augmented SFT artifacts. It is
+not the raw pipeline export. Each eligible source example contributes exactly
+one row: the original row is not included alongside additional copies.
+
+## Splits
+
+| Split | File | Rows |
+| --- | --- | ---: |
+| Train | `train.jsonl` | 12,571 |
+| Development/validation | `dev.jsonl` | 1,776 |
+| Test | `test.jsonl` | 3,574 |
+| **Total** | | **17,921** |
+
+The development file is exposed as the Hugging Face `validation` split while
+retaining the filename `dev.jsonl`.
+
+## Record format
+
+The main fields are:
+
+- `question_id`, `doc_id`: source identifiers.
+- `source_type`: `single_document` or `document_cluster`.
+- `question_family`: question-generation family.
+- `groundability`: `groundable`, `partially_groundable`, or `non_groundable`.
+- `system`: system message used for SFT.
+- `instruction`: question and evidence in one augmented presentation.
+- `response`: supervised assistant target.
+- `selected_answer`: selected final answer before chat formatting.
+- `cot_sample_index`: selected candidate index in the source pipeline.
+- `split`, `split_group`: leakage-controlled split metadata.
+- `cluster_id`: present for document-cluster examples.
+- `augmentation`: deterministic augmentation provenance, including the original
+  parent ID, template, evidence layout, evidence-ID scheme, shuffle flag, and
+  system-prompt variant.
+
+## Augmentation policy
+
+Instruction augmentation changes prompt wording, evidence presentation,
+evidence order, and citation-label style. It does not paraphrase the supervised
+answer. When the evidence-ID style changes, bracketed evidence citations in the
+target are remapped so they continue to refer to the correct evidence blocks.
+
+The augmentation seed is `42`. Exactly one augmented presentation is retained
+per eligible parent in train, validation, and test.
+
+## Quality and leakage controls
+
+The source split contained 18,743 rows. The SFT builder excluded 822 rows:
+
+| Reason | Train | Validation | Test | Total |
+| --- | ---: | ---: | ---: | ---: |
+| Empty evidence | 474 | 78 | 159 | 711 |
+| Incomplete target tags | 57 | 14 | 10 | 81 |
+| Invalid target citations or XML | 18 | 5 | 7 | 30 |
+
+Validation of this release found:
+
+- valid UTF-8 JSONL for every row;
+- one unique parent per row;
+- valid and ordered `<raciocinio>`/`<resposta>` target tags;
+- no response citations referring to unavailable evidence identifiers;
+- zero parent-sample overlap across train, validation, and test;
+- zero evidence-document overlap across train, validation, and test.
+
+See `dataset_manifest.json` for split distributions, exclusions, augmentation
+mix, and leakage-audit counts. See `SHA256SUMS` for release checksums.
+
+## Intended use
+
+Use `train` for supervised fine-tuning, `validation` for model selection and
+training diagnostics, and `test` only for final held-out evaluation. The model
+should be trained to answer exclusively from the supplied evidence, cite the
+evidence identifiers it uses, and explicitly state when the evidence is
+insufficient.
+
+## Limitations
+
+The examples are generated from scientific-document evidence and selected by
+an automated self-consistency pipeline. They may retain OCR artifacts, source
+document errors, or imperfect generated reasoning despite automated validation.
+The dataset should not be treated as a substitute for expert review in medical,
+engineering, or other high-stakes settings.
+

SHA256SUMS

@@ -0,0 +1,4 @@
+0089089414aa7d2c0eadbff38ed46f17a4f9da0d350ac193240eea25b332e190  train.jsonl
+e3463dd658de33e0683089a90320e1771de608bf5f50780f6fa62c436ed64aab  dev.jsonl
+532475e20b9184ac806dd2491a0f395eb06c1672f96e373a504d2d545ec8dbb7  test.jsonl
+540a43b543f37704bdfb85d59f6e01820cbfe40e80ca8d4c16fdb7f7a86ccdb5  dataset_manifest.json

dataset_manifest.json

@@ -0,0 +1,279 @@
+{
+  "total_included": 17921,
+  "splits": {
+    "train": {
+      "records": 12571,
+      "groups": 12571,
+      "source_type": {
+        "single_document": 10051,
+        "document_cluster": 2520
+      },
+      "question_family": {
+        "method_result_reasoning": 10051,
+        "cross_document_synthesis": 2520
+      },
+      "groundability": {
+        "partially_groundable": 7556,
+        "groundable": 3290,
+        "non_groundable": 1725
+      },
+      "sample_kind": {
+        "answer_cot": 12571
+      }
+    },
+    "validation": {
+      "records": 1776,
+      "groups": 1776,
+      "source_type": {
+        "single_document": 1426,
+        "document_cluster": 350
+      },
+      "question_family": {
+        "method_result_reasoning": 1426,
+        "cross_document_synthesis": 350
+      },
+      "groundability": {
+        "groundable": 479,
+        "non_groundable": 226,
+        "partially_groundable": 1071
+      },
+      "sample_kind": {
+        "answer_cot": 1776
+      }
+    },
+    "test": {
+      "records": 3574,
+      "groups": 3574,
+      "source_type": {
+        "single_document": 2854,
+        "document_cluster": 720
+      },
+      "question_family": {
+        "method_result_reasoning": 2854,
+        "cross_document_synthesis": 720
+      },
+      "groundability": {
+        "non_groundable": 479,
+        "partially_groundable": 2155,
+        "groundable": 940
+      },
+      "sample_kind": {
+        "answer_cot": 3574
+      }
+    }
+  },
+  "excluded": {
+    "train:empty_evidence": 474,
+    "train:incomplete_tags": 57,
+    "train:invalid_target_citations_or_xml": 18,
+    "validation:empty_evidence": 78,
+    "validation:incomplete_tags": 14,
+    "validation:invalid_target_citations_or_xml": 5,
+    "test:empty_evidence": 159,
+    "test:incomplete_tags": 10,
+    "test:invalid_target_citations_or_xml": 7
+  },
+  "group_overlap_count": {
+    "train__validation": 0,
+    "train__test": 0,
+    "validation__test": 0
+  },
+  "group_overlaps": {
+    "train__validation": [],
+    "train__test": [],
+    "validation__test": []
+  },
+  "input_mode": "presplit_leakage_controlled",
+  "input_paths": {
+    "train": "data\\OpenSciReason-PT-v1\\openscireason_pt.train.jsonl",
+    "validation": "data\\OpenSciReason-PT-v1\\openscireason_pt.dev.jsonl",
+    "test": "data\\OpenSciReason-PT-v1\\openscireason_pt.test.jsonl"
+  },
+  "input_records": {
+    "train": 13120,
+    "validation": 1873,
+    "test": 3750
+  },
+  "excluded_records": {
+    "train:empty_evidence": 474,
+    "train:incomplete_tags": 57,
+    "train:invalid_target_citations_or_xml": 18,
+    "validation:empty_evidence": 78,
+    "validation:incomplete_tags": 14,
+    "validation:invalid_target_citations_or_xml": 5,
+    "test:empty_evidence": 159,
+    "test:incomplete_tags": 10,
+    "test:invalid_target_citations_or_xml": 7
+  },
+  "output_paths": {
+    "train": "training\\sft\\openscireason_pt.augmented.sft.train.jsonl",
+    "validation": "training\\sft\\openscireason_pt.augmented.sft.validation.jsonl",
+    "test": "training\\sft\\openscireason_pt.augmented.sft.test.jsonl"
+  },
+  "seed": 42,
+  "evidence_document_overlap_count": {
+    "train__validation": 0,
+    "train__test": 0,
+    "validation__test": 0
+  },
+  "parent_sample_overlap_count": {
+    "train__validation": 0,
+    "train__test": 0,
+    "validation__test": 0
+  },
+  "instruction_augmentation": {
+    "enabled": true,
+    "strategy": "single_all",
+    "source_rows_replaced": true,
+    "rows_per_eligible_parent": 1,
+    "augmented_splits": [
+      "train",
+      "validation",
+      "test"
+    ],
+    "created_records": 17921,
+    "skipped_records": {},
+    "realized_mix": {
+      "train": {
+        "records": 12571,
+        "augmented": 12571,
+        "unique_parents": 12571,
+        "sft_template_id": {
+          "T1_minimal": 3796,
+          "T5_concise": 968,
+          "T3_evidence_first": 2455,
+          "T4_rag_style": 1830,
+          "T6_detailed": 940,
+          "T2_explicit_short": 2582
+        },
+        "evidence_layout": {
+          "original": 8878,
+          "source_block": 648,
+          "bullet": 1184,
+          "compact": 1861
+        },
+        "evidence_id_scheme": {
+          "source": 1195,
+          "evidence": 8841,
+          "s": 1239,
+          "ev": 1296
+        },
+        "evidence_shuffled": {
+          "False": 10118,
+          "True": 2453
+        },
+        "format_reminder": {
+          "False": 7482,
+          "True": 5089
+        },
+        "system_prompt_variant": {
+          "False": 11963,
+          "True": 608
+        }
+      },
+      "validation": {
+        "records": 1776,
+        "augmented": 1776,
+        "unique_parents": 1776,
+        "sft_template_id": {
+          "T2_explicit_short": 348,
+          "T1_minimal": 555,
+          "T5_concise": 119,
+          "T4_rag_style": 249,
+          "T3_evidence_first": 366,
+          "T6_detailed": 139
+        },
+        "evidence_layout": {
+          "original": 1259,
+          "bullet": 166,
+          "compact": 253,
+          "source_block": 98
+        },
+        "evidence_id_scheme": {
+          "s": 182,
+          "evidence": 1251,
+          "ev": 162,
+          "source": 181
+        },
+        "evidence_shuffled": {
+          "False": 1423,
+          "True": 353
+        },
+        "format_reminder": {
+          "True": 724,
+          "False": 1052
+        },
+        "system_prompt_variant": {
+          "False": 1693,
+          "True": 83
+        }
+      },
+      "test": {
+        "records": 3574,
+        "augmented": 3574,
+        "unique_parents": 3574,
+        "sft_template_id": {
+          "T4_rag_style": 496,
+          "T1_minimal": 1059,
+          "T2_explicit_short": 749,
+          "T3_evidence_first": 722,
+          "T5_concise": 284,
+          "T6_detailed": 264
+        },
+        "evidence_layout": {
+          "original": 2525,
+          "source_block": 167,
+          "compact": 559,
+          "bullet": 323
+        },
+        "evidence_id_scheme": {
+          "evidence": 2510,
+          "ev": 382,
+          "source": 342,
+          "s": 340
+        },
+        "evidence_shuffled": {
+          "False": 2869,
+          "True": 705
+        },
+        "format_reminder": {
+          "False": 2160,
+          "True": 1414
+        },
+        "system_prompt_variant": {
+          "True": 173,
+          "False": 3401
+        }
+      }
+    },
+    "response_policy": "unchanged_except_for_bracketed_evidence_id_remapping",
+    "target_final_mix": {
+      "sft_template_id": {
+        "T1_minimal": 0.3,
+        "T2_explicit_short": 0.2,
+        "T3_evidence_first": 0.2,
+        "T4_rag_style": 0.15,
+        "T5_concise": 0.075,
+        "T6_detailed": 0.075
+      },
+      "system_prompt": {
+        "main": 0.95,
+        "variant": 0.05
+      },
+      "evidence_layout": {
+        "original": 0.7,
+        "compact": 0.15,
+        "bullet": 0.1,
+        "source_block": 0.05
+      },
+      "evidence_id_scheme": {
+        "evidence": 0.7,
+        "ev": 0.1,
+        "source": 0.1,
+        "s": 0.1
+      },
+      "evidence_shuffled": 0.2,
+      "format_reminder": 0.4
+    }
+  }
+}

dev.jsonl

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e3463dd658de33e0683089a90320e1771de608bf5f50780f6fa62c436ed64aab
+size 37713225

test.jsonl

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:532475e20b9184ac806dd2491a0f395eb06c1672f96e373a504d2d545ec8dbb7
+size 75041449

train.jsonl

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0089089414aa7d2c0eadbff38ed46f17a4f9da0d350ac193240eea25b332e190
+size 266194815