Ompatil19/SupportLM-triage-dataset

SupportLM triage v1.0.2

Held-out customer support tickets with five-field triage labels, published so the SupportLM V1.0.2 adapter's scores are reproducible rather than self-reported.

Synthetic data. Every ticket was LLM-generated (llm_generated_v6_augmented) for this project. It has never seen real support traffic, and the synthetic-to-real gap is unmeasured.

Splits

Split	Rows	Purpose
train	4,000	the rows used for fine-tuning
validation	600	checkpoint selection
test	600	held out; all reported metrics come from here

Splits are disjoint by uid, verified at export time with a hard gate that refuses to stage if any overlap appears. Near-duplicate tickets cannot straddle the train/test boundary by construction.

Schema

{{
  "uid": "supportlm-001268",
  "split": "test",
  "domain": "healthcare",
  "company": "MediCare Plus",
  "category": "prescription",
  "intent": "dosage_question",
  "label_source": "offline",
  "user_text": "I'm experiencing severe anxiety ...",
  "labels": {{
    "emotion": "Anxious",
    "urgency": "Critical",
    "summary": "I require this refill immediately ...",
    "next_action": "CLINICAL_ESCALATION",
    "suggested_reply": "Ms. Dubois, I understand your distress ..."
  }}
}}

Field	Type	Notes
uid	string	stable ticket id; the split key
split	string	train / valid / test
domain	string	ecommerce, healthcare, finance, saas
company	string	fictional brand, e.g. ShopNova
category, intent	string	22 categories, 87 intents in the source corpus
label_source	string	offline (rule-based) or llm (Gemini-written)
user_text	string	the single customer message
labels.emotion	enum	Angry, Frustrated, Confused, Anxious, Neutral, Polite, Happy
labels.urgency	enum	Low, Medium, High, Critical
labels.summary	string	<=25 words, one sentence
labels.next_action	enum	1 of 42 fixed constants (PROCESS_REFUND, ESCALATE_HUMAN, ...)
labels.suggested_reply	string	reference reply, verbatim from the corpus - never generated by the model

Label provenance, per field

• emotion, urgency, next_action - rule-based from the corpus metadata, so they are deterministic and consistent.
• summary - 14 rows written by Gemini, the rest by a grounded extractive fallback, because free-tier API quota was exhausted mid-run. label_source marks which rows are which. This field mostly tests the summariser, not the model.
• suggested_reply - the corpus's own chosen_response, copied verbatim.

Loading

from datasets import load_dataset

ds = load_dataset("{DATASET_REPO}")
print(ds["test"][0]["labels"])

curl -LO https://huggingface.co/datasets/{DATASET_REPO}/resolve/main/data/test.jsonl

Benchmark

eval.yaml registers the test split as a benchmark task (triage_v1) so scores can be submitted through Hub Eval Results and appear on a leaderboard. It uses evaluation_framework: inspect-ai.

Baseline for comparison, measured on this exact split:

Metric	Base Qwen2.5-1.5B	+ SupportLM V1.0.2
Emotion accuracy	55.7%	99.0%
Urgency accuracy	19.8%	94.8%
Next-action accuracy	1.2%	56.3%
Reply ROUGE-L	0.141	0.735
Reply invents amounts	0.2%	58.8%

The last row is why this dataset is public: a reproducible benchmark should make the bad numbers as easy to find as the good ones.

Caveats for benchmark users

• next_action is imbalanced. 23.2% of the test split is a single action (EXPLAIN_PRODUCT), and urgency is 55.8% High. A majority-class baseline scores higher on next_action than the model does. Report macro-F1 or per-class numbers alongside accuracy.
• Reply fabrication rises with ROUGE-L. The tuned model invents amounts in 58.8% of replies versus 0.2% for the base model. ROUGE-L cannot distinguish a correct draft from a fluent invention, so score grounding separately.
• Synthetic, single-turn, English-only, one fictional brand per domain. It measures schema learning, not production fit.
• Not adversarially tested. Ticket text is untrusted input.

License

CC0-1.0 (public domain), matching the source corpus metadata.