kaushik-systalyze/customer-transcript-analytics

Customer Transcript Analytics

Curated customer-support and meeting transcripts mapped to a single fixed "analyze this transcript → compact JSON" prompt, for benchmarking batched offline LLM inference on realistic workloads.

Motivation and intended use

This dataset provides a realistic transcript-analytics workload for batched offline-inference experiments: throughput benchmarking and predicted-vs-observed throughput validation. Rows range from short support chats of a few hundred tokens to long multi-party meeting segments, and every row carries token accounting (input_tokens, expected_max_new_tokens, prefill/decode ratio estimates, length_bucket) so throughput can be analyzed by prompt length.

It is not a correctness benchmark for intent, sentiment, resolution, or summarization quality. Model outputs generated from these prompts should not be interpreted as validated labels.

Prompt template

Every row uses this exact template (the static prefix before the transcript is constant across rows; per-row static_prefix_tokens / static_prefix_reuse quantify it for prefix-caching analysis):

Analyze this customer-support transcript.

Return compact JSON with:
- intent
- sentiment
- resolved
- escalation_needed
- summary

Transcript:
{{transcript}}

expected_max_new_tokens is a documented heuristic, not a measurement: min(256, max(96, 96 + input_tokens // 64)). Token accounting is approximate and uses the Qwen/Qwen3.5-9B tokenizer (add_special_tokens=False).

Example usage (OpenAI-style request)

Each row's prompt is sent as a single user message, and the row's expected_max_new_tokens sets max_tokens. The model's JSON is generated at benchmark time — it is not stored in the dataset and must not be read as a validated label. The responses below are illustrative only.

Short customer-support row (short_control, ABCD, 189 input tokens):

{
  "model": "Qwen/Qwen3.5-9B",
  "temperature": 0,
  "max_tokens": 98,
  "messages": [
    {
      "role": "user",
      "content": "Analyze this customer-support transcript.\n\nReturn compact JSON with:\n- intent\n- sentiment\n- resolved\n- escalation_needed\n- summary\n\nTranscript:\nASSISTANT: Hello, how can i help you today\nUSER: I need to check on my order because the delivery time is wrong\n[... truncated for display ...]\nUSER: No, thank you\nASSISTANT: Okay, have a good day"
    }
  ]
}

Illustrative response:

{
  "intent": "delivery_time_inquiry",
  "sentiment": "neutral",
  "resolved": true,
  "escalation_needed": false,
  "summary": "Customer asked why their delivery time seemed wrong; agent confirmed the order is on schedule for standard 5-day shipping."
}

This particular row also carries gold metadata passed through from the source (target_json = {"intent": "status_delivery_time", ...}, has_gold_intent=true), which a future quality check could compare against — the throughput benchmark itself does not score outputs.

Long meeting segment (long_dialogue, AMI window, 4,805 input tokens, 4k_8k bucket): identical request shape, but with ~4.8k tokens of speaker-prefixed meeting transcript in the user message and max_tokens: 171. Such rows are prefill-dominated (prefill_ratio_estimate ≈ 0.97), in contrast to the short case above.

Notes: the 34-token static prefix is identical across all rows, so engines with prefix caching will reuse it (static_prefix_tokens / static_prefix_reuse quantify this per row). If your server supports response_format: {"type": "json_object"}, it keeps outputs compact, though the prompt already encourages that and correctness is not evaluated.

Splits

• source — Customer-relevant transcript-analytics prompts (demo believability). Mostly 512–4k-token task-oriented customer-service dialogues.
• long_dialogue — Longer transcript-like prompts for long-context throughput and throughput-prediction validation, covering 512 to 16k input tokens with most rows between 2k and 8k.
• short_control — Short realistic workloads (128–1024 input tokens) for benchmarking short-prompt serving behaviour.
• holdout_eval — Held out for predicted-vs-observed throughput validation ONLY. Not a quality-eval split. Do not use it to fit or select anything. Transcripts are disjoint from all other splits (normalized-hash check) and drawn from all source families and prompt lengths.

Source mix by split

• source (1566 rows): Taskmaster2: 544, SalesBot: 363, Taskmaster3: 331, MULTIWOZ2_2: 232, ABCD: 61, SGD: 35
• long_dialogue (1217 rows): QMSum: 716, ICSI: 273, AMI: 228
• short_control (3000 rows): ABCD: 545, SGD: 495, Taskmaster3: 488, Taskmaster2: 459, MULTIWOZ2_2: 434, bitext/Bitext-customer-support-llm-chatbot-training-dataset: 391, SalesBot: 188
• holdout_eval (1049 rows): QMSum: 200, Taskmaster2: 167, Taskmaster3: 150, MULTIWOZ2_2: 107, SGD: 93, ABCD: 88, SalesBot: 85, bitext/Bitext-customer-support-llm-chatbot-training-dataset: 61, ICSI: 59, AMI: 39

Token statistics

split	rows	mean in-tok	p50	p75	p90	p95	p99	mean exp-new	prefill	decode	prefix-reuse p90
source	1566	599.8	569	634	712	793	908	104.9	0.8492	0.1508	0.0654
long_dialogue	1217	3581.4	3045	5654	6542	7652	13680	150.4	0.9414	0.0586	0.0407
short_control	3000	304.0	300	384	441	472	504	100.3	0.7358	0.2642	0.2037
holdout_eval	1049	1163.5	520	863	3240	5443	7476	113.6	0.8224	0.1776	0.1576
overall	6832	1087.6	495	675	3233	5772	7709	112.3	0.8117	0.1883	0.1659

Length-bucket counts

split	short_under_512	512_1k	1k_2k	2k_4k	4k_8k	8k_16k
source	0	1560	6	0	0	0
long_dialogue	0	184	246	369	369	49
short_control	3000	0	0	0	0	0
holdout_eval	491	313	77	94	65	9

Length buckets on input_tokens: short_under_512 (<512), 512_1k, 1k_2k, 2k_4k, 4k_8k (4096–8192 inclusive), 8k_16k (8193–16384).

Licenses and redistribution

Each row carries license, redistribution_status, and source_attribution columns. Per-source terms:

source	config	license	redistribution status
Salesforce/dialogstudio	ABCD	MIT	redistributable
Salesforce/dialogstudio	Taskmaster2	CC BY 4.0	redistributable
Salesforce/dialogstudio	Taskmaster3	CC BY 4.0	redistributable
Salesforce/dialogstudio	SGD	CC BY-SA 4.0	redistributable (share-alike)
Salesforce/dialogstudio	SalesBot	none (no upstream license)	no-upstream-license; included by dataset owner decision (fully synthetic simulator/GPT-3 text)
Salesforce/dialogstudio	MULTIWOZ2_2	MIT	redistributable
Salesforce/dialogstudio	AMI	CC BY 4.0	redistributable
Salesforce/dialogstudio	ICSI	CC BY 4.0	redistributable
Salesforce/dialogstudio	QMSum	MIT	redistributable
bitext/Bitext-customer-support-llm-chatbot-training-dataset	—	CDLA-Sharing-1.0	redistributable (share-alike for data)

Notes: SGD-derived rows remain CC BY-SA 4.0 (share-alike); Bitext-derived rows remain CDLA-Sharing-1.0. SalesBot has no upstream license; it is included by explicit dataset-owner decision because its text is fully synthetic (user-simulator/GPT-3 generated) with no third-party human content — treat those rows accordingly if that risk profile does not fit your use.

Excluded sources

The following sources from the original target list were excluded (no raw rows included) because their terms do not clearly allow public redistribution:

• TweetSumm — Summary annotations are CC0, but the underlying tweet text comes from the Kaggle 'Customer Support on Twitter' (TWCS) corpus, licensed CC BY-NC-SA 4.0 (non-commercial). Upstream distributes only tweet IDs for this reason. (https://github.com/guyfe/Tweetsumm ; https://www.kaggle.com/datasets/thoughtvector/customer-support-on-twitter)
• ConvoSumm — Conflicting license signals (repo CC BY-SA 4.0 vs DialogStudio 'No License') and third-party content (NYT comments, Reddit, StackExchange) with unclear redistribution provenance; also off-domain for customer support. (https://github.com/Yale-LILY/ConvoSumm)
• MediaSum — Upstream restricts use to research purposes only (NPR/CNN copyrighted transcripts); not publicly redistributable. (https://github.com/zcgzcgzcg1/MediaSum (README usage restriction))
• DialogSum — Upstream README states CC BY-NC-SA 4.0 (non-commercial); sources include DailyDialog which carries the same restriction. (https://github.com/cylnlp/dialogsum)
• SAMSum — CC BY-NC-ND 4.0 (non-commercial, no derivatives). (https://huggingface.co/datasets/knkarthick/samsum (card: cc-by-nc-nd-4.0))
• tuanhqv123/meeting-summarization — Mixed-source dataset whose card admits roughly half is MeetingBank (CC BY-NC 4.0, non-commercial) and lists license 'other'. (https://huggingface.co/datasets/tuanhqv123/meeting-summarization)
• monodox/customer-support-and-service — License is fine (Apache-2.0) but the dataset contains only 8 nested records — too trivial to contribute. (https://huggingface.co/datasets/monodox/customer-support-and-service)

Construction

• Built from raw Salesforce/dialogstudio JSON files (configs listed in metadata.json → source_files_used; for large task-oriented configs a deterministic subset of dialogue files was used) plus the Bitext CSV.
• Task-oriented dialogues are flattened from the DialogStudio log field into USER: / ASSISTANT: turns; meeting transcripts come from original dialog info → dialog history as SPEAKER: utterance lines.
• Long-meeting segmentation: AMI/ICSI/QMSum meetings run 9k–18k+ tokens, so rows in the 1k–8k buckets are contiguous, non-overlapping segments of a single meeting (never concatenations of unrelated transcripts). QMSum is cut on its own annotated topic spans; AMI/ICSI use utterance-aligned windows. A meeting contributes either windows or one whole-meeting row (≤16k), never both, to avoid shared long prefixes that would skew prefix-cache behaviour.
• Rows over 16384 input tokens were rejected (never truncated) and are counted per source in source_report.json. Deduplication is by normalized transcript hash; holdout_eval transcripts are disjoint from all other splits.
• Built by a deterministic fixed-seed pipeline (seed 42).

Label policy

• Existing source labels are preserved when trivially available (reference_intent from ABCD subflow / SalesBot intent / Bitext intent; reference_summary from AMI abstractive summaries and QMSum query answers).
• No labels are synthesized — nothing is LLM-generated, heuristically inferred, or guessed from text. has_gold_* flags are true only for fields that came directly from source data.
• No source provides sentiment, resolution, or escalation labels, so reference_sentiment, reference_resolved, reference_escalation_needed are null everywhere and the corresponding has_gold_* flags are false.
• target_json is assembled only from available source fields and may be partial or null.
• The benchmark runs the model on prompt and generates fresh outputs; the reference fields are optional metadata for future quality checks only.

Limitations

• Several sources are synthetic or proxy data (SalesBot and Bitext are fully synthetic; Taskmaster/SGD/MultiWOZ are Wizard-of-Oz or simulated task-oriented dialogues) — they look like customer service but are not real support calls.
• Meeting transcripts (AMI/ICSI/QMSum) are research-meeting recordings, not customer service; they provide realistic long-transcript shape, not customer-support content.
• Long rows are contiguous segments of longer meetings (see Construction).
• long_dialogue is modest in size because most long-form sources (MediaSum, MeetingBank, SAMSum, DialogSum) failed license review.
• Intended for inference throughput experiments, not model quality evaluation.

Files

• data/<split>-*.parquet — the four splits
• metadata.json — build configuration (seed, tokenizer, template, files used)
• token_stats.json — per-split token statistics
• source_report.json — per-source row accounting, license evidence, exclusions