Upload European defect repair SFT dataset v0.1

.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
+data/all.jsonl filter=lfs diff=lfs merge=lfs -text
+data/train.jsonl filter=lfs diff=lfs merge=lfs -text
+metadata/all_metadata.jsonl filter=lfs diff=lfs merge=lfs -text

README.md

@@ -0,0 +1,67 @@
+---
+dataset_info:
+  pretty_name: OpenEuroLLM Defect Repair SFT v1
+  license: cc-by-sa-4.0
+---
+
+# oellm-eu-defect-repair-sft-v1
+
+Monolingual SFT repair data for European-language generation defects observed
+after Qwen 2B/4B/9B post-training. This dataset is designed to repair
+degeneration, repetition loops, short answers, morphology damage, and
+Bulgarian/Russian language leakage.
+
+## Strict SFT Schema
+
+Each row in `data/*.jsonl` uses exactly:
+
+```json
+{"messages":[{"role":"user","content":"..."},{"role":"assistant","content":"..."}],"lang":"is"}
+```
+
+Provenance, source URL, license, source id, quality statistics, and hashes are
+kept in `metadata/all_metadata.jsonl`, `source_registry.json`,
+`metadata/quality_report.json`, and `manifest.json`.
+
+## Source Policy
+
+Version `v0.1.0` is built from native Wikipedia text via
+`wikimedia/wikipedia` 2023-11-01 snapshots. Rows are not raw machine
+translations. Prompt templates are short target-language instructions and
+assistant turns are human-authored native article passages.
+
+License: Wikimedia text is redistributed under CC-BY-SA terms, with attribution
+and share-alike requirements. Keep the sidecar provenance files when using or
+redistributing filtered subsets.
+
+## Counts
+
+- `bg`: 2000
+- `da`: 3000
+- `et`: 8000
+- `fi`: 1000
+- `ga`: 8000
+- `hr`: 3000
+- `hu`: 3000
+- `is`: 8000
+- `lt`: 3000
+- `lv`: 3000
+- `mt`: 7352
+- `pl`: 1000
+- `ro`: 1000
+- `sk`: 3000
+- `sl`: 3000
+
+Total strict SFT rows: `57352`.
+
+## Intended Use
+
+Use as an upweighted repair shard in SFT:
+
+- Tier 1 (`is`, `ga`, `mt`, `et`): high upweight for fluency and answer length.
+- Tier 2 (`hr`, `sl`, `lt`, `lv`, `da`, `hu`, `sk`): medium upweight for loop repair.
+- `bg`: anti-Russian leakage shard with Russian-marker filters.
+- Tier 4 (`ro`, `pl`, `fi`): light touch-up.
+
+Run `scripts/validate_defect_repair_sft.py --dataset-dir data/defect_repair/oellm-eu-defect-repair-sft-v1` before
+packing to parquet or mixing into SFT.

data/all.jsonl

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f00728ebb48932cb51bdb3eab3977cbd1afd608f8d27bb3cb400e8576e347c06
+size 124077794

data/train.jsonl

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:bcb7394f8aac55831c829df6f182391305d64577fe868a632a49bae4fde5873e
+size 121712262

manifest.json

@@ -0,0 +1,60 @@
+{
+  "dataset_id": "oellm-eu-defect-repair-sft-v1",
+  "version": "v0.1.0",
+  "schema": "chat_messages_strict_sft",
+  "created_by": "build_defect_repair_sft.py",
+  "total_rows": 57352,
+  "counts_by_split": {
+    "train": 56239,
+    "validation": 1113
+  },
+  "counts_by_language": {
+    "bg": 2000,
+    "da": 3000,
+    "et": 8000,
+    "fi": 1000,
+    "ga": 8000,
+    "hr": 3000,
+    "hu": 3000,
+    "is": 8000,
+    "lt": 3000,
+    "lv": 3000,
+    "mt": 7352,
+    "pl": 1000,
+    "ro": 1000,
+    "sk": 3000,
+    "sl": 3000
+  },
+  "target_counts_by_language": {
+    "is": 8000,
+    "ga": 8000,
+    "mt": 8000,
+    "et": 8000,
+    "hr": 3000,
+    "sl": 3000,
+    "lt": 3000,
+    "lv": 3000,
+    "da": 3000,
+    "hu": 3000,
+    "sk": 3000,
+    "bg": 2000,
+    "ro": 1000,
+    "pl": 1000,
+    "fi": 1000
+  },
+  "source_ids": [
+    "wikimedia_wikipedia_20231101"
+  ],
+  "license_categories": [
+    "sharealike"
+  ],
+  "strict_export_files": [
+    "data/train.jsonl",
+    "data/validation.jsonl",
+    "data/all.jsonl"
+  ],
+  "metadata_files": [
+    "metadata/all_metadata.jsonl",
+    "metadata/quality_report.json"
+  ]
+}

metadata/all_metadata.jsonl

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:7da2e2d1dff27d17f952ea7e90c35b3b2eb00a2525d0d7fc5ce5988b41d7e09d
+size 50798912

metadata/quality_report.json

@@ -0,0 +1,228 @@
+{
+  "dataset_id": "oellm-eu-defect-repair-sft-v1",
+  "version": "v0.1.0",
+  "parameters": {
+    "min_assistant_words": 80,
+    "max_assistant_words": 400,
+    "validation_fraction": 0.02,
+    "target_scale": 1.0,
+    "source_scan_multiplier": 18,
+    "max_source_rows_per_language": null
+  },
+  "accepted_by_language": {
+    "bg": 2000,
+    "da": 3000,
+    "et": 8000,
+    "fi": 1000,
+    "ga": 8000,
+    "hr": 3000,
+    "hu": 3000,
+    "is": 8000,
+    "lt": 3000,
+    "lv": 3000,
+    "mt": 7352,
+    "pl": 1000,
+    "ro": 1000,
+    "sk": 3000,
+    "sl": 3000
+  },
+  "seen_articles_by_language": {
+    "bg": 1232,
+    "da": 3603,
+    "et": 12445,
+    "fi": 706,
+    "ga": 14202,
+    "hr": 3542,
+    "hu": 2287,
+    "is": 14809,
+    "lt": 3042,
+    "lv": 3463,
+    "mt": 5743,
+    "pl": 878,
+    "ro": 994,
+    "sk": 3197,
+    "sl": 4162
+  },
+  "rejections_by_language": {
+    "bg": {
+      "assistant_too_long": 31,
+      "bg_russian_marker": 13,
+      "no_usable_chunk": 60,
+      "repeated_4gram": 67,
+      "script_ratio_low": 8
+    },
+    "da": {
+      "assistant_too_long": 1,
+      "bad_title": 27,
+      "missing_native_marker_chars": 1,
+      "no_usable_chunk": 1441,
+      "repeated_4gram": 41
+    },
+    "et": {
+      "assistant_too_long": 13,
+      "bad_title": 16,
+      "missing_native_marker_chars": 5,
+      "no_usable_chunk": 6287,
+      "repeated_4gram": 235,
+      "script_ratio_low": 7
+    },
+    "fi": {
+      "no_usable_chunk": 100,
+      "repeated_4gram": 7
+    },
+    "ga": {
+      "assistant_too_long": 15,
+      "bad_title": 14,
+      "no_usable_chunk": 7723,
+      "repeated_4gram": 175
+    },
+    "hr": {
+      "assistant_too_long": 6,
+      "no_usable_chunk": 1481,
+      "repeated_4gram": 40
+    },
+    "hu": {
+      "assistant_too_long": 4,
+      "bad_title": 5,
+      "no_usable_chunk": 389,
+      "repeated_4gram": 77
+    },
+    "is": {
+      "assistant_too_long": 5,
+      "bad_title": 44,
+      "no_usable_chunk": 8491,
+      "repeated_4gram": 134,
+      "script_ratio_low": 1
+    },
+    "lt": {
+      "assistant_too_long": 39,
+      "no_usable_chunk": 796,
+      "repeated_4gram": 419,
+      "script_ratio_low": 1
+    },
+    "lv": {
+      "assistant_too_long": 3,
+      "no_usable_chunk": 1206,
+      "repeated_4gram": 69,
+      "script_ratio_low": 1
+    },
+    "mt": {
+      "assistant_too_long": 9,
+      "bad_title": 5,
+      "duplicate_answer": 6,
+      "missing_native_marker_chars": 9,
+      "no_usable_chunk": 1401,
+      "repeated_4gram": 509,
+      "script_ratio_low": 4
+    },
+    "pl": {
+      "assistant_too_long": 1,
+      "no_usable_chunk": 215,
+      "repeated_4gram": 29
+    },
+    "ro": {
+      "assistant_too_long": 1,
+      "bad_title": 1,
+      "no_usable_chunk": 328,
+      "repeated_4gram": 62
+    },
+    "sk": {
+      "assistant_too_long": 7,
+      "no_usable_chunk": 1199,
+      "repeated_4gram": 54,
+      "script_ratio_low": 1
+    },
+    "sl": {
+      "assistant_too_long": 21,
+      "bad_title": 9,
+      "missing_native_marker_chars": 1,
+      "no_usable_chunk": 2098,
+      "repeated_4gram": 47
+    }
+  },
+  "assistant_word_stats_by_language": {
+    "is": {
+      "min": 80,
+      "max": 400,
+      "mean": 244.5
+    },
+    "ga": {
+      "min": 80,
+      "max": 400,
+      "mean": 241.89
+    },
+    "mt": {
+      "min": 80,
+      "max": 400,
+      "mean": 290.21
+    },
+    "et": {
+      "min": 80,
+      "max": 400,
+      "mean": 267.78
+    },
+    "hr": {
+      "min": 80,
+      "max": 400,
+      "mean": 279.63
+    },
+    "sl": {
+      "min": 80,
+      "max": 400,
+      "mean": 273.7
+    },
+    "lt": {
+      "min": 80,
+      "max": 400,
+      "mean": 283.19
+    },
+    "lv": {
+      "min": 80,
+      "max": 400,
+      "mean": 269.66
+    },
+    "da": {
+      "min": 80,
+      "max": 400,
+      "mean": 274.98
+    },
+    "hu": {
+      "min": 80,
+      "max": 400,
+      "mean": 298.02
+    },
+    "sk": {
+      "min": 80,
+      "max": 400,
+      "mean": 285.24
+    },
+    "bg": {
+      "min": 80,
+      "max": 400,
+      "mean": 316.27
+    },
+    "ro": {
+      "min": 80,
+      "max": 400,
+      "mean": 299.79
+    },
+    "pl": {
+      "min": 80,
+      "max": 400,
+      "mean": 294.56
+    },
+    "fi": {
+      "min": 80,
+      "max": 400,
+      "mean": 322.23
+    }
+  },
+  "hard_filters": [
+    "assistant_80_to_400_words_by_default",
+    "no_4gram_repeated_more_than_3_times",
+    "target_script_ratio",
+    "language_marker_char_presence",
+    "bg_russian_marker_ban",
+    "url_ban"
+  ]
+}

source_registry.json

@@ -0,0 +1,16 @@
+{
+  "wikimedia_wikipedia_20231101": {
+    "source_id": "wikimedia_wikipedia_20231101",
+    "name": "Wikimedia Wikipedia 2023-11-01 dumps",
+    "source_url": "https://huggingface.co/datasets/wikimedia/wikipedia",
+    "source_license": "CC-BY-SA-4.0 / GFDL under Wikimedia Terms of Use; verify imported-page attribution when needed",
+    "license_id": "cc-by-sa-4.0",
+    "license_category": "sharealike",
+    "license_filter_tags": [
+      "attribution_required",
+      "sharealike",
+      "native_wikipedia"
+    ],
+    "redistribution_status": "redistributable_sharealike"
+  }
+}