kaushik-systalyze/customer-transcript-holdout-eval

Customer Transcript Holdout Eval

Curated customer-support and transcript-analytics prompts 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 carry 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; per-row static_prefix_tokens / static_prefix_reuse quantify the constant prefix 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. Responses are illustrative only.

Notes: the 34-token static prefix is identical across all rows, so engines with prefix caching will reuse it. If your server supports response_format: {"type": "json_object"}, it keeps outputs compact, though correctness is not evaluated.

Split

• 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.

Source mix

• 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
holdout_eval	1049	1163.5	520	863	3240	5443	7476	113.6	0.8224	0.1776	0.1576

Length-bucket counts

split	short_under_512	512_1k	1k_2k	2k_4k	4k_8k	8k_16k
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 represented in this dataset:

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

Notes: SGD-derived rows remain CC BY-SA 4.0 (share-alike); Bitext-derived rows remain CDLA-Sharing-1.0 when present. SalesBot has no upstream license when present; 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.

Excluded sources

The following sources from the original target list were excluded 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 plus the Bitext CSV where represented in this split.
• 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 rows are contiguous, non-overlapping meeting segments or whole-meeting rows, never concatenations of unrelated transcripts.
• Rows over 16384 input tokens were rejected and never truncated. Deduplication is by normalized transcript hash.
• 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.
• 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; reference fields are optional metadata for future quality checks only.

Limitations

• Some sources are synthetic or proxy data when represented (SalesBot and Bitext are fully synthetic; Taskmaster/SGD/MultiWOZ are Wizard-of-Oz or simulated task-oriented dialogues).
• Meeting transcripts, when represented, are research-meeting recordings; they provide realistic long-transcript shape, not customer-support content.
• Intended for inference throughput experiments, not model quality evaluation.

Files

• data/holdout_eval-*.parquet - the holdout_eval split
• metadata.json - build configuration for this split
• token_stats.json - token statistics for this split
• source_report.json - source row accounting, license evidence, and exclusions for this split