julia-1-swe-data

The exact training and evaluation data for 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 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:

{"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:

{"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
• Code: niko0xdev/julia-1-zorch
• Related dataset with the same label schema (not used to train julia-1-swe): niko0xdev/zorch-triage-dataset

Citation

@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}}
}