Add julia-1-swe training and evaluation data with dataset card

.gitattributes

@@ -58,3 +58,10 @@ 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/r120/train.jsonl filter=lfs diff=lfs merge=lfs -text
+data/r122/train.jsonl filter=lfs diff=lfs merge=lfs -text
+data/r126/train.jsonl filter=lfs diff=lfs merge=lfs -text
+data/r132/train.jsonl filter=lfs diff=lfs merge=lfs -text
+data/r133/train.jsonl filter=lfs diff=lfs merge=lfs -text
+data/r134/train.jsonl filter=lfs diff=lfs merge=lfs -text
+data/train_union/train.jsonl filter=lfs diff=lfs merge=lfs -text

README.md

@@ -0,0 +1,247 @@
+---
+license: apache-2.0
+task_categories:
+- text-classification
+language:
+- vi
+- ko
+pretty_name: julia-1-swe training and evaluation data
+size_categories:
+- 10K<n<100K
+tags:
+- synthetic
+- ticket-triage
+- sdlc
+- multilingual
+- zorch
+configs:
+- config_name: default
+  data_files:
+  - {split: train, path: data/train_union/train.jsonl}
+  - {split: validation, path: data/train_union/validation.jsonl}
+- config_name: eval_tasks
+  data_files:
+  - {split: r118, path: data/eval_tasks/r118.jsonl}
+  - {split: r119, path: data/eval_tasks/r119.jsonl}
+  - {split: r121, path: data/eval_tasks/r121.jsonl}
+- config_name: r118
+  data_files:
+  - {split: train, path: data/r118/train.jsonl}
+  - {split: validation, path: data/r118/validation.jsonl}
+  - {split: validation_hard, path: data/r118/validation_hard.jsonl}
+- config_name: r119
+  data_files:
+  - {split: validation, path: data/r119/validation.jsonl}
+  - {split: validation_hard, path: data/r119/validation_hard.jsonl}
+- config_name: r120
+  data_files:
+  - {split: train, path: data/r120/train.jsonl}
+  - {split: augmentation, path: data/r120/augmentation.jsonl}
+- config_name: r121
+  data_files:
+  - {split: validation, path: data/r121/validation.jsonl}
+  - {split: validation_hard, path: data/r121/validation_hard.jsonl}
+- config_name: r122
+  data_files:
+  - {split: train, path: data/r122/train.jsonl}
+  - {split: augmentation, path: data/r122/augmentation.jsonl}
+- config_name: r126
+  data_files:
+  - {split: train, path: data/r126/train.jsonl}
+  - {split: augmentation, path: data/r126/augmentation.jsonl}
+- config_name: r132
+  data_files:
+  - {split: train, path: data/r132/train.jsonl}
+  - {split: augmentation, path: data/r132/augmentation.jsonl}
+- config_name: r133
+  data_files:
+  - {split: train, path: data/r133/train.jsonl}
+  - {split: augmentation, path: data/r133/augmentation.jsonl}
+- config_name: r134
+  data_files:
+  - {split: train, path: data/r134/train.jsonl}
+  - {split: augmentation, path: data/r134/augmentation.jsonl}
+---
+
+# julia-1-swe-data
+
+The exact training and evaluation data for [niko0xdev/julia-1-swe](https://huggingface.co/niko0xdev/julia-1-swe), a Zorch SDLC ticket triage model with three classification heads: **role**, **action** and **risk**. All data is synthetic: generated Vietnamese (with and without diacritics) and Korean ticket titles. No real tickets were used.
+
+The rows come from the `data/v3/noisy_vi_ko_compositional_r118` to `r134` rounds of the [julia-1-zorch](https://github.com/niko0xdev/julia-1-zorch) repository.
+
+## What was used for julia-1-swe
+
+| Config / split | Rows | Used for julia-1-swe |
+|---|---|---|
+| `default` / `train` | 63,324 | **Training** (R134 `train.jsonl` + R126 `augmentation.jsonl`, deduplicated) |
+| `default` / `validation` | 3,564 | **Validation and checkpoint selection** (R118 + R119 + R121 `validation_hard`, 1,188 tasks x 3 heads) |
+| `eval_tasks` / `r118`, `r119`, `r121` | 396 each | **Reported evaluation** (task-level, exact match over all three heads) |
+
+The training union was built with `data/train_union/build_latest_union.py` (copied from `scripts/linux/build_latest_union.py`). It concatenates R134 `train.jsonl` (50,670 rows) and R126 `augmentation.jsonl` (12,672 rows), drops exact duplicates on (state, question, target), and removes any row whose normalized title appears in the evaluation sets (0 were removed). Result: 63,324 rows from 21,108 tickets.
+
+## Per-round configs
+
+Each round is also provided on its own. "Contained in training" was checked by exact match of (state, question, target) against `default/train`.
+
+| Config | Split | Rows | Contained in julia-1-swe training | Description |
+|---|---|---|---|---|
+| `r118` | train | 9,504 | Yes (100% of unique rows) | Base compositional training split (1,584 tasks, two title forms each) |
+| `r118` | validation | 1,188 | No (evaluation) | 396 held-out label combinations |
+| `r118` | validation_hard | 1,188 | No (evaluation) | Same rows as `validation`, used by the trainer interface |
+| `r119` | validation | 1,188 | No (evaluation) | Holdout |
+| `r119` | validation_hard | 1,188 | No (evaluation) | Same rows as `validation` |
+| `r120` | train | 23,760 | Yes (100%) | R118 train plus compact, reordered, noisy augmentation |
+| `r120` | augmentation | 14,256 | Yes (100%) | R120 augmentation rows only |
+| `r121` | validation | 1,188 | No (evaluation) | Holdout with an explicit risk cue in every title |
+| `r121` | validation_hard | 1,188 | No (evaluation) | Same rows as `validation` |
+| `r122` | train | 37,998 | Yes (100%) | R120 train plus action-focused short titles |
+| `r122` | augmentation | 14,238 | Yes (100%) | R122 augmentation rows only |
+| `r126` | train | 50,670 | Yes (100%) | R122 train plus informal no-diacritic Vietnamese |
+| `r126` | augmentation | 12,672 | **Yes, directly** | R126 augmentation rows only (part of the union) |
+| `r132` | train | 41,166 | Yes (100%) | R122 train plus Korean `업무:` context titles |
+| `r132` | augmentation | 3,168 | Yes (100%) | R132 augmentation rows only |
+| `r133` | train | 47,502 | Yes (100%) | R122 train plus explicit-risk titles in holdout style |
+| `r133` | augmentation | 9,504 | Yes (100%) | R133 augmentation rows only |
+| `r134` | train | 50,670 | **Yes, directly** | R122 train plus R132 and R133 augmentations (part of the union) |
+| `r134` | augmentation | 12,672 | Yes (100%) | R132 and R133 augmentation rows only |
+
+The training rounds are cumulative, so every training or augmentation file from R118 to R134 is a subset of `default/train`. None of the R118, R119 or R121 validation rows appear in training. Each round folder also has its `stats.json` and, where available, the generator notes as `NOTES.md`. Rounds R123 to R125 and R127 to R131 are not in the repository.
+
+## Data structure
+
+### Head rows (`default` and per-round configs)
+
+One row per ticket per head:
+
+```json
+{"meta": {"decision": "role", "language": "vi", "scenario_id": "r118-vi-product-implement-normal-v0",
+          "source": "synthetic_compositional_r118", "style": "shared_vocabulary_train", "variant": "paraphrase_0"},
+ "options": ["product", "business_analysis", "research", "design", "frontend", "mobile", "backend", "data", "ai", "qa", "devops", "security"],
+ "question": "Which role should own this task?",
+ "state": {"title": "làm mới quy tắc gói cước cho khách hàng mới"},
+ "target": 0, "type": "choice"}
+```
+
+| Field | Description |
+|---|---|
+| `state.title` | Ticket title (the only input field in this data) |
+| `question` | Head question: role, action or risk (see Labels) |
+| `options` | Label list for the head |
+| `target` | Index of the correct label in `options` |
+| `type` | Always `choice` |
+| `meta.decision` | Head: `role`, `action` or `risk` |
+| `meta.language` | `vi`, `vi_nodiac` (Vietnamese without diacritics) or `ko` |
+| `meta.scenario_id` | Generated scenario id (encodes round, language and labels) |
+| `meta.source` | Generator round |
+| `meta.style`, `meta.variant` | Title style and paraphrase variant |
+
+### Task rows (`eval_tasks`)
+
+One row per ticket with all three expected labels:
+
+```json
+{"expected": {"action": "implement", "risk": "security", "role": "product"},
+ "meta": {"lang": "vi", "source": "synthetic_compositional_r118", "style": "shared_vocabulary_holdout"},
+ "task": {"id": "r118-vi-product-implement-security", "title": "quy tắc gói cước cho khách hàng mới: làm mới bảo vệ tài khoản"}}
+```
+
+## Labels
+
+| Head | Question | Labels |
+|---|---|---|
+| role (12) | Which role should own this task? | product, business_analysis, research, design, frontend, mobile, backend, data, ai, qa, devops, security |
+| action (11) | What kind of work is this task? | implement, document, fix, test, research, review, analyze, investigate, design, deploy, refactor |
+| risk (5) | What is the main risk category of this task? | normal, production, security, data_sensitive, architecture |
+
+### Label counts (tickets per label, `default` config)
+
+| role | train | validation |
+|---|---|---|
+| devops | 1,848 | 81 |
+| design | 1,840 | 69 |
+| business_analysis | 1,814 | 87 |
+| qa | 1,808 | 93 |
+| mobile | 1,778 | 96 |
+| data | 1,760 | 99 |
+| research | 1,757 | 96 |
+| ai | 1,745 | 111 |
+| backend | 1,729 | 102 |
+| product | 1,713 | 111 |
+| security | 1,665 | 123 |
+| frontend | 1,651 | 120 |
+
+| action | train | validation |
+|---|---|---|
+| fix | 2,086 | 69 |
+| design | 2,044 | 72 |
+| refactor | 1,956 | 102 |
+| deploy | 1,952 | 105 |
+| test | 1,914 | 111 |
+| investigate | 1,900 | 111 |
+| implement | 1,868 | 120 |
+| research | 1,867 | 117 |
+| analyze | 1,862 | 123 |
+| review | 1,860 | 126 |
+| document | 1,799 | 132 |
+
+| risk | train | validation |
+|---|---|---|
+| production | 4,266 | 234 |
+| architecture | 4,226 | 243 |
+| security | 4,212 | 237 |
+| data_sensitive | 4,208 | 240 |
+| normal | 4,196 | 234 |
+
+### Language split (head rows)
+
+| Language | train | validation |
+|---|---|---|
+| vi_nodiac (Vietnamese without diacritics) | 28,494 | 1,188 |
+| ko | 18,990 | 1,188 |
+| vi | 15,840 | 1,188 |
+
+Each `eval_tasks` split has 132 tasks per language.
+
+### Styles in `default`
+
+Train: compact_noisy_composition 14,256; action_focused_short 14,235; no_diacritic_short_noisy 12,672; explicit_risk_holdout_style 9,504; shared_vocabulary_train 9,489; shared_korean_work_context 3,168. Validation: shared_vocabulary_holdout (R118), reordered_shared_vocabulary (R119), explicit_risk_holdout (R121), 1,188 rows each.
+
+## How the data was created
+
+- **R118** generates titles from shared role, action and risk phrase banks in three language forms, then holds out 20% of the label combinations per language for validation (seed 20261118). Every phrase cue used in validation also occurs in training.
+- **R119** is a separate holdout in a reordered shared-vocabulary style (seed 20261119).
+- **R121** is a holdout on the combinations reserved by R118, with titles not present in the R120 training file. Every title carries an explicit risk cue (seed 20261121).
+- **R120, R122, R126, R132, R133, R134** are training-only augmentations that use only R118 training combinations: compact and noisy titles (R120), action-focused short titles (R122), informal no-diacritic Vietnamese with abbreviations and typos (R126), Korean `업무:` context (R132), explicit-risk titles in the R121 separator style (R133), and R132 + R133 together (R134).
+- Per the generator notes, no real tickets or real test examples were read during generation.
+- Of the 315 label combinations in the evaluation sets, 308 also occur in the training union, according to `data/train_union/stats.json`.
+
+## Intended use
+
+- Reproducing or extending the julia-1-swe fine-tune.
+- Benchmarking small multi-head triage classifiers on Vietnamese and Korean ticket titles.
+- Not a measure of accuracy on real tickets.
+
+## Limitations
+
+- **Synthetic only.** Titles are generated from phrase banks. They are short, title-only inputs with no descriptions or acceptance criteria.
+- **Train and evaluation come from the same generator family.** The evaluation sets share vocabulary and phrase cues with training and hold out label combinations, not vocabulary. Scores measure in-distribution compositional generalization. 308 of the 315 evaluation label combinations also appear in training.
+- **Small evaluation sets.** 396 tasks per set, 1,188 in total.
+- **No English.** Despite the base model supporting English, this data contains only Vietnamese and Korean.
+- **Cumulative rounds.** Per-round training files overlap heavily. Do not combine them without deduplicating.
+
+## Related
+
+- Model: [niko0xdev/julia-1-swe](https://huggingface.co/niko0xdev/julia-1-swe)
+- Code: [niko0xdev/julia-1-zorch](https://github.com/niko0xdev/julia-1-zorch)
+- Related dataset with the same label schema (not used to train julia-1-swe): [niko0xdev/zorch-triage-dataset](https://huggingface.co/datasets/niko0xdev/zorch-triage-dataset)
+
+## Citation
+
+```bibtex
+@misc{julia1swedata2026,
+  title        = {julia-1-swe-data: Synthetic Vietnamese and Korean SDLC Ticket Triage Data},
+  author       = {niko0xdev},
+  year         = {2026},
+  howpublished = {\url{https://huggingface.co/datasets/niko0xdev/julia-1-swe-data}}
+}
+```

data/r118/NOTES.md

@@ -0,0 +1,13 @@
+# R118 synthetic compositional dataset
+
+This round uses generated Vietnamese, no-diacritic Vietnamese, and Korean task titles. The training and validation splits share the same role, action, and risk phrase banks, while holding out combinations of those labels. No real tickets or real test examples were read.
+
+## Files
+
+- `train.jsonl`: 9,504 classification rows from 1,584 generated tasks, with two title forms per task.
+- `validation.jsonl`: 1,188 classification rows from 396 held-out combinations.
+- `validation_hard.jsonl`: copy of the held-out classification rows for the existing hard-validation interface.
+- `validation_tasks_hard.jsonl`: task-level inputs and expected labels for exact-match evaluation.
+- `stats.json`: generation seed, split counts, and disjointness checks.
+
+The split uses seed `20261118` and holds out 20% of the combinations in each language. The generator checks that every phrase cue used in validation also occurs in training. The validation set measures compositional generalisation within a shared synthetic vocabulary; it does not measure accuracy on real tickets.

data/r118/stats.json

@@ -0,0 +1,34 @@
+{
+  "dataset_kind": "synthetic short-title Vietnamese, no-diacritic Vietnamese, and Korean paraphrases",
+  "real_tickets_used": false,
+  "real_test_read": false,
+  "seed": 20261118,
+  "validation_fraction": 0.2,
+  "train_task_ids": 1584,
+  "validation_task_ids": 396,
+  "train_rows": 9504,
+  "validation_rows": 1188,
+  "task_ids_disjoint": true,
+  "validation_phrase_cues_seen_in_training": true,
+  "validation_phrase_cue_count": 164,
+  "train_languages": {
+    "vi": 528,
+    "vi_nodiac": 528,
+    "ko": 528
+  },
+  "validation_languages": {
+    "vi": 132,
+    "vi_nodiac": 132,
+    "ko": 132
+  },
+  "train_decisions": {
+    "role": 3168,
+    "action": 3168,
+    "risk": 3168
+  },
+  "validation_decisions": {
+    "role": 396,
+    "action": 396,
+    "risk": 396
+  }
+}

data/r119/stats.json

@@ -0,0 +1,22 @@
+{
+  "dataset_kind": "synthetic reordered-title holdout for the R118 train split",
+  "real_tickets_used": false,
+  "real_test_read": false,
+  "source_training_split": "R118 train.jsonl",
+  "source_holdout_split": "R118 validation_tasks_hard.jsonl",
+  "seed": 20261119,
+  "validation_tasks": 396,
+  "validation_rows": 1188,
+  "heldout_combinations_disjoint_from_r118_train": true,
+  "titles_disjoint_from_r118_train_and_validation": true,
+  "validation_languages": {
+    "vi": 132,
+    "vi_nodiac": 132,
+    "ko": 132
+  },
+  "validation_decisions": {
+    "role": 396,
+    "action": 396,
+    "risk": 396
+  }
+}

data/r120/NOTES.md

@@ -0,0 +1,7 @@
+# R120 synthetic training augmentation
+
+This training-only augmentation adds compact, reordered, and noisy Vietnamese, no-diacritic Vietnamese, and Korean titles. It derives labels only from R118 training combinations. R118 validation and the 329 real tasks are not read for generation. Use this file alongside the existing R118 training split; evaluate on a separate holdout.
+
+- `train.jsonl`: R118 training rows plus generated augmentation rows
+- `augmentation.jsonl`: generated rows only
+- `stats.json`: counts and split provenance

data/r120/stats.json

@@ -0,0 +1,29 @@
+{
+  "dataset_kind": "synthetic compact-title augmentation from R118 training combinations",
+  "real_tickets_used": false,
+  "real_test_read": false,
+  "source_training_split": "data/v3/noisy_vi_ko_compositional_r118/train.jsonl",
+  "seed": 20261120,
+  "source_scenarios": 1584,
+  "augmentation_tasks": 4752,
+  "base_training_rows": 9504,
+  "augmentation_rows": 14256,
+  "training_rows": 23760,
+  "task_combinations_disjoint_from_r118_validation": true,
+  "titles_unique_within_language": true,
+  "languages": {
+    "ko": 1584,
+    "vi": 1584,
+    "vi_nodiac": 1584
+  },
+  "decisions": {
+    "role": 7920,
+    "action": 7920,
+    "risk": 7920
+  },
+  "styles": {
+    "compact_reordered": 1,
+    "short_separators": 1,
+    "korean_compacted": 1
+  }
+}

data/r120/train.jsonl

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b36a91c2a869d480ca3b63190b6287cbf83362bdc2c693be904020228a70e1b6
+size 12380712

data/r121/NOTES.md

@@ -0,0 +1,3 @@
+# R121 synthetic holdout
+
+This holdout uses combinations reserved by R118 and titles not present in the R120 training file. Every title includes an explicit risk cue, including a low-risk cue for normal tasks, so the label is observable from the input. No real tickets or real test examples were read.

data/r121/stats.json

@@ -0,0 +1,16 @@
+{
+  "dataset_kind": "synthetic held-out combinations with explicit risk context",
+  "real_tickets_used": false,
+  "real_test_read": false,
+  "source_holdout": "data/v3/noisy_vi_ko_compositional_r118/validation_tasks_hard.jsonl",
+  "seed": 20261121,
+  "validation_tasks": 396,
+  "validation_rows": 1188,
+  "task_combinations_disjoint_from_r120_train": true,
+  "titles_disjoint_from_r120_train": true,
+  "languages": {
+    "vi": 132,
+    "vi_nodiac": 132,
+    "ko": 132
+  }
+}

data/r122/NOTES.md

@@ -0,0 +1,7 @@
+# R122 synthetic action-focused training data
+
+R122 adds three short action-focused title variants for each R118 training combination and includes the R120 training file. It balances Vietnamese, no-diacritic Vietnamese, and Korean, and keeps the R118 held-out combinations out of training. No real tickets or real test examples were read.
+
+- `train.jsonl`: R120 training rows plus R122 augmentation rows
+- `augmentation.jsonl`: R122 rows only
+- `stats.json`: counts and generation provenance

data/r122/stats.json

@@ -0,0 +1,25 @@
+{
+  "dataset_kind": "synthetic action-focused augmentation from R118 training combinations",
+  "real_tickets_used": false,
+  "real_test_read": false,
+  "base_training_file": "data/v3/noisy_vi_ko_compositional_r120/train.jsonl",
+  "source_training_split": "data/v3/noisy_vi_ko_compositional_r118/train.jsonl",
+  "seed": 20261122,
+  "source_scenarios": 1584,
+  "augmentation_tasks": 4746,
+  "skipped_base_title_duplicates": 6,
+  "base_training_rows": 23760,
+  "augmentation_rows": 14238,
+  "training_rows": 37998,
+  "titles_disjoint_from_base_training_file": true,
+  "languages": {
+    "ko": 1578,
+    "vi": 1584,
+    "vi_nodiac": 1584
+  },
+  "decisions": {
+    "role": 4746,
+    "action": 4746,
+    "risk": 4746
+  }
+}

data/r122/train.jsonl

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4a628b69d068c4ec65573a53e75af3a90bb986978f74d6b5e4a3ac3f7d22ffbc
+size 19804697

data/r126/NOTES.md

@@ -0,0 +1,7 @@
+# R126 synthetic no-diacritic Vietnamese training data
+
+R126 adds short, informal, no-diacritic Vietnamese variants with abbreviations, mixed English terms, and typo noise. All source label combinations come from the R118 training split; held-out combinations and real tickets were not read.
+
+- `train.jsonl`: R122 training rows plus R126 augmentation rows
+- `augmentation.jsonl`: R126 rows only
+- `stats.json`: counts and generation provenance

data/r126/stats.json

@@ -0,0 +1,23 @@
+{
+  "dataset_kind": "synthetic no-diacritic Vietnamese short-title augmentation from R118 training combinations",
+  "real_tickets_used": false,
+  "real_test_read": false,
+  "base_training_file": "data/v3/noisy_vi_ko_compositional_r122/train.jsonl",
+  "source_training_split": "data/v3/noisy_vi_ko_compositional_r118/train.jsonl",
+  "source_language": "vi_nodiac",
+  "seed": 20261206,
+  "source_scenarios": 528,
+  "augmentation_tasks": 4224,
+  "augmentation_rows": 12672,
+  "base_training_rows": 37998,
+  "training_rows": 50670,
+  "titles_disjoint_from_base_training_file": true,
+  "languages": {
+    "vi_nodiac": 12672
+  },
+  "decisions": {
+    "role": 4224,
+    "action": 4224,
+    "risk": 4224
+  }
+}

data/r126/train.jsonl

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d07867d2bf420e24d00235ae7fb3715fa4df10b1287df02702e649953f9f2c73
+size 26260084

data/r132/NOTES.md

@@ -0,0 +1,7 @@
+# R132 synthetic Korean training data
+
+R132 adds two Korean title variants per R118 training combination, using the shared `업무:` context seen in validation. It only uses combinations from the R118 training split; no held-out combinations or real tickets were read.
+
+- `train.jsonl`: R122 training rows plus R132 augmentation rows
+- `augmentation.jsonl`: R132 rows only
+- `stats.json`: counts and generation provenance

data/r132/stats.json

@@ -0,0 +1,23 @@
+{
+  "dataset_kind": "synthetic Korean shared-context augmentation from R118 training combinations",
+  "real_tickets_used": false,
+  "real_test_read": false,
+  "base_training_file": "data/v3/noisy_vi_ko_compositional_r122/train.jsonl",
+  "source_training_split": "data/v3/noisy_vi_ko_compositional_r118/train.jsonl",
+  "seed": 20261212,
+  "source_scenarios": 528,
+  "variants_per_scenario": 2,
+  "augmentation_tasks": 1056,
+  "augmentation_rows": 3168,
+  "base_training_rows": 37998,
+  "training_rows": 41166,
+  "titles_disjoint_from_base_training_file": true,
+  "languages": {
+    "ko": 3168
+  },
+  "decisions": {
+    "role": 1056,
+    "action": 1056,
+    "risk": 1056
+  }
+}

data/r132/train.jsonl

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:2cb9c17d6bfebdffa2884029915a656bfb29c85f452f22bc2066e96aaea2149e
+size 21484812

data/r133/NOTES.md

@@ -0,0 +1,7 @@
+# R133 synthetic training data in holdout title style
+
+R133 adds two explicit-risk title variants for each R118 training combination, using shuffled role, action, and risk cues with the same separator pattern used by R121. Only R118 training combinations are used; held-out combinations and real tickets were not read.
+
+- `train.jsonl`: R122 training rows plus R133 augmentation rows
+- `augmentation.jsonl`: R133 rows only
+- `stats.json`: counts and generation provenance

data/r133/stats.json

@@ -0,0 +1,25 @@
+{
+  "dataset_kind": "synthetic training augmentation in the R121 title style",
+  "real_tickets_used": false,
+  "real_test_read": false,
+  "base_training_file": "data/v3/noisy_vi_ko_compositional_r122/train.jsonl",
+  "source_training_split": "data/v3/noisy_vi_ko_compositional_r118/train.jsonl",
+  "seed": 20261213,
+  "source_scenarios": 1584,
+  "variants_per_scenario": 2,
+  "augmentation_tasks": 3168,
+  "augmentation_rows": 9504,
+  "base_training_rows": 37998,
+  "training_rows": 47502,
+  "titles_disjoint_from_base_training_file": true,
+  "languages": {
+    "ko": 1056,
+    "vi": 1056,
+    "vi_nodiac": 1056
+  },
+  "decisions": {
+    "role": 3168,
+    "action": 3168,
+    "risk": 3168
+  }
+}

data/r133/train.jsonl

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:bed7dace9c59118514073fb1d7b06d90dfbac4bfcd1a4d9c89788d39aff7cf65
+size 24938432

data/r134/NOTES.md

@@ -0,0 +1,7 @@
+# R134 synthetic mixed-context training data
+
+R134 combines R132 Korean `업무:` context variants and R133 shuffled role, action, and risk titles. Both augmentations use only R118 training combinations; no held-out combinations or real tickets were read.
+
+- `train.jsonl`: R122 training rows plus R132 and R133 augmentations
+- `augmentation.jsonl`: R132 and R133 rows only
+- `stats.json`: counts and generation provenance

data/r134/stats.json

@@ -0,0 +1,24 @@
+{
+  "dataset_kind": "synthetic mixed-context training augmentation",
+  "real_tickets_used": false,
+  "real_test_read": false,
+  "base_training_file": "data/v3/noisy_vi_ko_compositional_r122/train.jsonl",
+  "korean_augmentation_file": "data/v3/noisy_vi_ko_compositional_r132/augmentation.jsonl",
+  "holdout_style_augmentation_file": "data/v3/noisy_vi_ko_compositional_r133/augmentation.jsonl",
+  "base_training_rows": 37998,
+  "korean_augmentation_rows": 3168,
+  "holdout_style_augmentation_rows": 9504,
+  "augmentation_rows": 12672,
+  "training_rows": 50670,
+  "titles_unique_by_language": true,
+  "languages": {
+    "vi": 15840,
+    "vi_nodiac": 15840,
+    "ko": 18990
+  },
+  "decisions": {
+    "role": 16890,
+    "action": 16890,
+    "risk": 16890
+  }
+}

data/r134/train.jsonl

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c28987ac3fc53ff8fb3d8ba7b7fc1cf6bddf70ed096f295fc2c44a05a576c6c3
+size 26618547

data/train_union/build_latest_union.py

@@ -0,0 +1,31 @@
+"""Build data/v3/colab_latest_union: R134 train + R126 no-diacritic augmentation (deduped),
+combined R118/R119/R121 eval files, and a leakage report (titles + label combos)."""
+import json, unicodedata
+from pathlib import Path
+V = Path("data/v3"); OUT = V / "colab_latest_union"; OUT.mkdir(exist_ok=True)
+rd = lambda p: [json.loads(l) for l in Path(p).read_text(encoding="utf-8").splitlines() if l.strip()]
+norm = lambda s: " ".join(unicodedata.normalize("NFC", s or "").lower().split())
+EVAL = {s: V / f"noisy_vi_ko_compositional_{s}" for s in ("r118", "r119", "r121")}
+ev_titles, ev_combos, tasks, rows = set(), set(), [], []
+for s, d in EVAL.items():
+    t = rd(d / "validation_tasks_hard.jsonl"); tasks += t; rows += rd(d / "validation_hard.jsonl")
+    for x in t:
+        ev_titles.add(norm(x["task"]["title"])); e = x["expected"]; ev_combos.add((e["role"], e["action"], e["risk"]))
+train, seen = [], set()
+for p in (V / "noisy_vi_ko_compositional_r134/train.jsonl", V / "noisy_vi_ko_compositional_r126/augmentation.jsonl"):
+    for r in rd(p):
+        k = json.dumps([r["state"], r["question"], r["target"]], sort_keys=True, ensure_ascii=False)
+        if k in seen: continue
+        seen.add(k); train.append(r)
+leak = [r for r in train if norm(r["state"].get("title")) in ev_titles]
+train = [r for r in train if norm(r["state"].get("title")) not in ev_titles]
+# label-combo overlap among train scenarios (scenario = same title)
+by_title = {}
+for r in train: by_title.setdefault(norm(r["state"].get("title")), {})[r["meta"].get("decision")] = r["options"][r["target"]]
+tr_combos = {(h.get("role"), h.get("action"), h.get("risk")) for h in by_title.values() if len(h) == 3}
+w = lambda n, xs: (OUT / n).write_text("".join(json.dumps(x, ensure_ascii=False) + "\n" for x in xs), encoding="utf-8")
+w("train.jsonl", train); w("validation_tasks_hard.jsonl", tasks); w("validation_hard.jsonl", rows)
+st = {"train_rows": len(train), "train_scenarios": len(by_title), "removed_title_leaks": len(leak),
+      "eval_tasks": len(tasks), "eval_rows": len(rows), "eval_combos": len(ev_combos),
+      "eval_combos_seen_in_train": len(ev_combos & tr_combos)}
+(OUT / "stats.json").write_text(json.dumps(st, indent=2)); print(st)

data/train_union/stats.json

@@ -0,0 +1,9 @@
+{
+  "train_rows": 63324,
+  "train_scenarios": 21108,
+  "removed_title_leaks": 0,
+  "eval_tasks": 1188,
+  "eval_rows": 3564,
+  "eval_combos": 315,
+  "eval_combos_seen_in_train": 308
+}

data/train_union/train.jsonl

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c30621970f05f908a857f8f9b80c8bdba8e2db4bc8621941faa2acfe7e4e0e9e
+size 33065514