Swift-Support/swift-support-tickets-1.0

Swift Support Tickets (v1.0)

Banking support tickets in five language tracks — English, Sinhala, Tamil, Singlish (romanized Sinhala) and Tanglish (romanized Tamil) — each labeled for three triage tasks: intent (77-way), sentiment (binary) and priority (3-way).

Derived from BANKING77. This is the training data behind Swift-Support/labse-intent-1.0 and Swift-Support/labse-priority-1.0.

Loading

from datasets import load_dataset

ds = load_dataset("Swift-Support/swift-support-tickets-1.0")              # all five, pooled
si = load_dataset("Swift-Support/swift-support-tickets-1.0", "sinhala")   # one track

Configs: all (default), english, sinhala, singlish, tamil, tamilish.

Structure

field	type	description
id	int	Ticket id. The same id is the same ticket in every language.
language	string	english, sinhala, singlish, tamil or tamilish
text	string	The ticket in this track's language — the model input
text_en	string	The original English text, for alignment and inspection
category	string	77-way BANKING77 intent (the column is named category, not intent)
sentiment	string	Neutral or Negative
priority	string	Low, Medium or High

Splits

split	tickets	rows (× 5 languages)
train	8,500	42,500
validation	1,498	7,490
test	3,079	15,395

The split is drawn once on id, then fanned out to all five languages (split identity e7b5934392cd; validation is carved out of the official BANKING77 train file, stratified on the 77-way intent; test is the official BANKING77 test file, untouched).

ℹ️ Why the viewer reports ~131k rows. The corpus is 65,385 unique rows. It ships twice: once as the pooled all configuration (the split table above) and once as five per-language configurations of 13,077 rows each. The viewer sums every configuration, so it shows 130,770. Load all or the per-language configs — never both, or every row appears twice.

⚠️ Do not re-split these rows randomly. The same ticket appears five times under one id. Splitting on rows instead of on id puts the English copy of a ticket in train and its Sinhala copy in test, and every score after that is measuring memorisation. If you need a different split, draw it on id and fan it out.

Label provenance — read before using the labels as ground truth

The three label columns do not have the same standing:

label	source	agreement with human annotation
category (intent)	inherited from BANKING77 — real human labels	— (ground truth)
sentiment	LLM prompt, v8 (2026-08-19)	negative-F1 0.7931, agreement 97.6%, κ = 0.780
priority	LLM prompt	macro-F1 0.7722, agreement 80.4%, κ = 0.644

Sentiment and priority were generated by an LLM prompt and then benchmarked against a 500-ticket subset annotated by hand. The figures above are that benchmark: the labeling rule itself agrees with a human roughly 77–80% of the time. A model that scores 0.89 against these priority labels has learned the rule, not human judgement. Report both numbers.

Sentiment went through one relabeling pass (v5 → v8, 2026-08-19): the earlier v5 prompt agreed with human annotation at only 0.5769 negative-F1 (κ = 0.55) on the same 500-ticket set; v8 replaced 711 of 13,077 sentiment labels and lifted that to 0.7931. The sentiment column in this dataset already holds v8 labels — the 0.5769 figure describes a prompt version no longer shipped here and should not be quoted against this data.

Corrected 2026-09-08. This card previously gave the v8 ceiling as 0.7812 (agreement 97.2%, κ = 0.766). That figure scored the prompt's raw output, not the labels actually shipped in this dataset — the two differ on 8 of the 500 benchmark tickets. Against the shipped sentiment column the agreement is 0.7931 negative-F1 (95% CI [0.6667, 0.8956]; raw agreement 0.976, Cohen's κ = 0.780). Quote 0.7931 when the subject is this data. The priority figure below was always correct, because priority was never relabelled and its shipped column and prompt output are identical.

The interval is wide because the benchmark holds only 31 Negative tickets, annotated by one person. Treat 0.7931 as a point estimate with real uncertainty, not a hard bound.

Only category carries real human ground truth.

Class balance

Sentiment is heavily skewed — roughly 94% Neutral — so accuracy is a misleading metric on it; the Swift project scores Negative-F1 instead. Priority runs about 55% Low / 36% Medium / 9% High, and is scored with macro-F1.

Known issues

These are measured properties of the data, published rather than papered over.

1. The romanized tracks are synthetic. Singlish is rule-generated from the Sinhala text and Tanglish is machine-translated — neither is human-typed. They are cleaner and more regular than real code-mixed writing (measured Singlish out-of-vocabulary rate is 1.18%, identical to English's). Any conclusion about romanized or code-mixed handling drawn from this data is an optimistic upper bound and should be re-measured on human-typed text before it is trusted.

2. Train/test text overlap, larger in the translated tracks.

track	test texts also in train	% of unique test texts
english	6	0.19%
sinhala	71	2.31%
singlish	84	2.73%
tamil	70	2.31%
tamilish	7	0.23%

The English overlap is inherited from BANKING77's own train/test files — distinct tickets that happen to share wording. The translated tracks add more because translation collapses distinct English questions onto one target string ("Where can I use my card?" and "At which ATMs can I use this card?" can land on the same Sinhala sentence). Ticket ids never cross splits; this is text-level, not id-level, leakage.

3. One cross-language label misalignment. Ticket id 5160 ("When will the transaction go through?") carries category = transfer_not_received_by_recipient in the tamilish track and pending_card_payment in the other four. Sentiment and priority are aligned for every ticket. 1 row of 9,998; left as-is so the published data matches the source.

4. Translation quality is unaudited at scale. Sinhala and Tamil text is machine-translated from English and was not reviewed ticket by ticket.

5. ⚠️ The tamilish test rendering does not match its train rendering. Tamilish test text is 60.3% out-of-vocabulary against tamilish train; every other track is 20–30%, and tamilish dev is an ordinary 22.3%. Tamilish test text also averages more characters but fewer words than tamilish train (71.5 / 8.74 vs 68.6 / 9.79), and tamilish is the only track with duplicate test strings (3,034 unique of 3,079).

The tamilish rendering of the BANKING77 test file was evidently produced by a different transliteration process from the rendering of the train file. Known issue 2 above records the same fact from the other side: tamilish has the lowest train/test text overlap of any track (0.23%), which reads as "less leakage" but is really a provenance split.

Consequence: any model score on the tamilish test split is contaminated by this mismatch and is a lower bound. It hits fine-grained tasks hardest — 77-way intent drops ~22 points from dev to test on this track, while 3-way priority drops ~11. Dev-side tamilish numbers are unaffected. This is scheduled for correction in v1.1; until then, do not report tamilish test scores as a measure of romanized-Tamil handling.

Licence & attribution

Derived from BANKING77 (PolyAI), released under CC-BY-4.0, and redistributed under the same licence. Translations, romanizations and the sentiment/priority labels are additions by the Swift project.

See Citation below for the full BANKING77 entry and how to cite this work.

---

Citation

Citing BANKING77 alone is not sufficient attribution for this corpus. BANKING77 supplies the English text and the 77-way intent labels — 20% of the rows and one of the three label columns. The rest is this project's work:

Rows whose text is new (Sinhala, Tamil, Singlish, Tanglish)	52,308 of 65,385 — 80%
Rows inherited from BANKING77 (English)	13,077 — 20%
Label columns inherited	1 (category / intent)
Label columns added	2 (sentiment, priority) — 26,154 new per-ticket assignments
Split	the frozen id-level split is this project's, not BANKING77's

If you use the translated tracks, the sentiment or priority labels, the frozen split, or any of the models, cite this work. If you additionally use the English text or the intent labels, cite BANKING77 as well. Both licences are CC-BY-4.0, so attribution is a licence condition, not a courtesy.

This work — the corpus

@misc{swift_tickets_2026,
  author       = {Sithija Seneviratne and Ruththiragayan Sutharsan and Shazan Shaheed},
  title        = {Swift Support Tickets: A Multilingual Ticket-Triage Corpus for Sinhala,
                  Tamil and English, Including Romanized Code-Mixed Text},
  year         = {2026},
  publisher    = {Hugging Face},
  howpublished = {\url{https://huggingface.co/datasets/Swift-Support/swift-support-tickets-1.0}},
  note         = {Derived from BANKING77. Adds four language tracks and the sentiment and
                  priority label columns. Paper in preparation.}
}

This work — the models

@misc{swift_models_2026,
  author       = {Sithija Seneviratne and Ruththiragayan Sutharsan and Shazan Shaheed},
  title        = {Swift: Multilingual Baselines for Support-Ticket Triage in Sinhala,
                  Tamil and English},
  year         = {2026},
  publisher    = {Hugging Face},
  howpublished = {\url{https://huggingface.co/Swift-Support}},
  note         = {Intent, sentiment and priority classifiers. Paper in preparation.}
}

BANKING77 — the source of the English text and the intent labels

@inproceedings{casanueva2020banking77,
  author    = {Casanueva, I{\~n}igo and Tem{\v c}inas, Tadas and Gerz, Daniela and
               Henderson, Matthew and Vuli{\'c}, Ivan},
  title     = {Efficient Intent Detection with Dual Sentence Encoders},
  booktitle = {Proceedings of the 2nd Workshop on Natural Language Processing for
               Conversational AI},
  pages     = {38--45},
  year      = {2020},
  publisher = {Association for Computational Linguistics},
  doi       = {10.18653/v1/2020.nlp4convai-1.5},
  url       = {https://aclanthology.org/2020.nlp4convai-1.5/}
}