The Call Playbook Dataset contains annotated samples from real enterprise sales conversations across 5 binary classification tasks, each capturing a critical signal in the B2B sales process. It was constructed from 50 real sales calls ranging from 30 to 90 minutes in length, and was released alongside the ACL Findings 2026 paper "Distilling Examples into Task Instructions: Enhanced In-Context Learning for Real-World B2B Conversations."

In B2B sales, automatically classifying conversation segments at scale across diverse and evolving intents requires handling scarce labeled data, keeping annotation overhead low, and avoiding per-intent fine-tuning. This dataset was built to study exactly that setting. Each task is small, realistic, and grounded in genuine enterprise sales dialogue, making it well suited for few-shot in-context learning (ICL), prompt-based classification, and knowledge-extraction research.

🎯 Tasks

Each task is a binary classification problem over a conversation snippet:

Task	Positive class definition	Example positive
Business Goals	Desired outcomes or strategic objectives articulated by the prospect	"We need to cut churn by 20% before Q3."
Decision Criteria	Specific attributes, features, or evaluation metrics used by the prospect to assess potential solutions	"Security compliance is non-negotiable for us."
Decision Makers	Individuals or roles identified as having authority or influence over the purchasing decision	"This ultimately goes to our VP of Finance."
Decision Making Process	The series of steps or procedures described by the prospect for arriving at a final decision	"We run a 3-week POC with two vendors in parallel."
Pain Points	Inefficiencies, obstacles, or needs expressed by the prospect that they aim to address through a potential solution	"Our current tool breaks on calls longer than an hour."

📊 Dataset Statistics

Task	Train Samples	Train Calls	Test Samples	Test Calls	Avg Words
Business Goals	200	25	200	25	~284
Decision Criteria	200	25	200	25	~276
Decision Makers	200	25	200	25	~267
Decision Making Process	200	20	200	21	~267
Pain Points	200	25	200	25	~286

Each task is a binary classification problem (label 1 = positive, 0 = negative).

🧱 Data Structure

Fields

Each task contains train.csv / test.csv with the following columns:

Column	Type	Description
`id`	int	Unique identifier within the split
`text`	string	Speaker-tagged conversation snippet
`label`	int	`1` if the concept is present, `0` otherwise

Example rows

id,text,label
0,"[PROSPECT_A] We need to reduce churn before Q3. [SELLER_A] Absolutely, let's talk about how.",1
1,"[SELLER_A] Great, I'll send over the proposal tonight. [PROSPECT_A] Sounds good.",0

Conversations are represented as speaker-tagged utterance sequences, preserving turn structure while enabling flexible windowing.

Speaker Tags

Speaker roles are annotated inline within each snippet:

Tag	Role
`[PROSPECT_A]`, `[PROSPECT_B]`	Prospect-side speakers
`[SELLER_A]`, `[SELLER_B]`	Seller-side speakers
`[SPEAKER_A]`, ...	Generic speaker (role unknown)

🚀 Usage

Load any task by passing its name as the configuration:

from datasets import load_dataset

# Load a single task
ds = load_dataset("gong-io-research/call-playbook", "business_goals")

print(ds["train"][0])
# {'id': 0, 'text': '[PROSPECT_A] We need to reduce churn ...', 'label': 1}

# Available configs:
# business_goals, decision_criteria, decision_makers,
# decision_making_process, pain_points

Iterate over all tasks:

from datasets import load_dataset

tasks = [
    "business_goals",
    "decision_criteria",
    "decision_makers",
    "decision_making_process",
    "pain_points",
]

for task in tasks:
    ds = load_dataset("gong-io-research/call-playbook", task)
    print(task, ds["train"].num_rows, ds["test"].num_rows)

🔒 Privacy & Anonymization

All data has been rigorously anonymized before release. Named entities were identified and systematically replaced with fictional alternatives that preserve conversational realism:

Entity type	Replacements	Sample substitutions
Organizations	120+	"Quantum Solutions", "Zenith Innovations", "Nebula Technologies"
Persons	130+	"Alex", "Jordan", "Casey", "Taylor"
Products	80+	"CodeCraft", "QuantaQuery", "NebulaNet"
Locations	180+	"Varthevia", "Brindmere", "Corswick"
URLs / IDs / Phones / Emails	20+	anonymized placeholder formats

No real personal data, customer identities, or proprietary information remains in the released dataset.

The complete entity replacement table is available as part of the dataset in replacements.json at the repository root.

🧪 Intended Uses

Recommended for:

Few-shot and zero-shot in-context learning (ICL) research
Prompt-based and instruction-based text classification
Knowledge-extraction methods (e.g., distilling examples into criteria or task descriptions)
Benchmarking LLMs on long, real-world conversational inputs
B2B / sales conversation understanding

Out of scope / limitations:

The dataset is English-only and drawn from enterprise sales calls. It may not generalize to other domains or languages.
Because named entities have been replaced with fictional substitutes, any statistics about the specific organizations, people, or products mentioned in the text do not reflect real-world distributions.
Each task is relatively small by design, reflecting the low-resource setting it was built to study.

📎 Associated Code

The full experimental framework, including our knowledge extraction methods (Criteria-Ex, Description-Ex, and iterative variants), is available on GitHub:

👉 github.com/gong-io/call-playbook

📚 Citation

If this dataset is useful to your research, please cite:

📄 Paper: aclanthology.org/2026.findings-acl.1631 · arXiv:2606.15641

@inproceedings{rotman-etal-2026-distilling,
    title = "Distilling Examples into Task Instructions: Enhanced In-Context Learning for Real-World {B}2{B} Conversations",
    author = "Rotman, Guy  and
      Kopilov, Adi  and
      Zalmanson, Danit Berger  and
      Allouche, Omri",
    editor = "Liakata, Maria  and
      Moreira, Viviane P.  and
      Zhang, Jiajun  and
      Jurgens, David",
    booktitle = "Findings of the {A}ssociation for {C}omputational {L}inguistics: {ACL} 2026",
    month = jul,
    year = "2026",
    address = "San Diego, California, United States",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2026.findings-acl.1631/",
    pages = "32590--32613",
    ISBN = "979-8-89176-395-1",
}

📄 License

This dataset is released under the Gong License. See the LICENSE file for the full terms and conditions governing use, redistribution, and permitted purposes.

Built by Gong Research · Gong License

Downloads last month: 55

Total file size:

3.03 MB

Paper for gong-io-research/call-playbook

Distilling Examples into Task Instructions: Enhanced In-Context Learning for Real-World B2B Conversations

Paper • 2606.15641 • Published Jun 14 • 5