76,000 Hours of Labeled Real Conversation Transcripts (Text Only)

Data Product Features

• Multi-turn, long context: Whole calls, not snippets. Good for training and testing long-conversation memory.
• Labeled intent on every turn: flow_role and behavior show why each turn was said, not just what.
• Built-in reward signal: behavior_effects scores each coaching technique by the listener's emotional change. Use it for preference data or reward modeling.
• Persona and role data: Clear coach, student and sales roles for role-play and agent fine-tuning.
• Group session structure: Hot seats, Q&A and teaching blocks, with the student in focus.
• Stages, summary and themes: Each call has a summary, a main theme, supporting themes and titled stages.
• Tone on every segment: A tone label plus a strength of low, medium or high.
• Key moments and emotional arcs: Insights, breakthroughs, resistance, and each speaker's path from -3 to +3.
• Voice measurements as numbers: Speaking speed, pitch, loudness and pauses are kept in the JSON, even with no audio.
• Clean privacy layer: Names, businesses, places, dates and amounts are replaced with numbered placeholders.

Distribution

• Size: About 21 GB.
• Data Volume: About 76,000 hours; 80,350 recordings; about 16.7 million transcript segments; 11 fields per segment, plus call-level data.

Usage

• Supervised fine-tuning: Coaching, sales and support agents.
• Preference and reward data: From the technique-to-outcome scores.
• Agent role-play: With real coach, student and sales roles.
• Dialogue evaluation: Test models on long, real conversations.
• Summarization and intent classification: With ready-made summaries and turn labels.

Coverage

• Geographic Coverage: Global; most participants are in the United States, Canada, the United Kingdom and Australia.
• Time Range: 2015 - 2026.
• Demographics: Adult entrepreneurs and small-business owners, talking with expert business coaches, a sales team and colleagues. Language: English.

License

AI Training Rights

• Use the Data Product to train, fine-tune, and evaluate machine learning models, including large language models.
• Incorporate Data Product content into models and commercialize resulting model outputs.
• Create derivative works (model weights, embeddings, etc.) for any lawful purpose.

• The Data Product itself may not be sold, redistributed, or shared outside of licensed usage.
• Licensee may not attempt to identify any person in the data.
• Licensee must comply with all applicable laws, including data protection and privacy regulations.

Who Can Use It

• LLM labs: For post-training and alignment.
• Agent startups: For coaching, sales or support agents.
• Alignment and safety researchers: For studies of supportive dialogue.
• Data scientists: For intent, emotion and summarization models.

Data Dictionary

| Column Name | Data Type | Description | Possible Values/Notes |
|-------------|-----------|-------------|-----------------------|
| call_id | string | Random ID for the call | e.g. CC-WYKCVAP5 |
| collection | string | Which collection the call belongs to | coaching_calls, sales_calls, team_calls |
| call_type | string | Kind of session | e.g. coaching_session |
| duration | number | Call length in seconds | e.g. 4289.34 |
| recording_period.year, .quarter | integer | When the call was recorded | Year and quarter only, for privacy |
| language | string | Language code | en |
| audio_quality.snr_db, .clipping, .channels | number | Signal-to-noise ratio, clipping share and channel count of the original recording | |
| source_format | object | Original recording format | Container, codec, bitrate, sample rate, channels |
| asr_engine, asr_mean_confidence | string, number | Speech recognizer used and its average confidence | Confidence from 0 to 1 |
| participants[] | array | Everyone who spoke: placeholder, pseudonym, role, how the role was assigned, and confidence | Roles: coach, student, participant |
| primary_theme, secondary_themes | string, array | Main and supporting topics | e.g. mindset_confidence, offer_design, launch_planning |
| summary | string | Short summary of the whole call | |
| stages[] | array | Titled stages of the call, with start and end times, segment range and summary | |
| blocks[] | array | Session formats within the call, with the student in focus | hot_seat, qa, teaching_block |
| segments[].id | integer | Order of the segment in the call | |
| segments[].start, .end | number | Start and end time, in seconds | |
| segments[].speaker | string | Placeholder for the speaker | e.g. [Coach A], [Student S-9831] |
| segments[].text | string | What was said, with identifying details replaced | e.g. [Person 3], [Business 2], [Product 1], [Location 1] |
| segments[].tones[] | array | Tone and strength | Tone e.g. encouraging, frustrated, curious, empathetic, realization, relieved; strength: low, medium, high |
| segments[].flow_role | string | What the segment does in the conversation | e.g. check_in, exploration, teaching, problem_statement, reframe, commitment, close |
| segments[].behavior | string | Coaching technique used, if any | e.g. diagnostic_question, normalize, give_example, reframe, celebrate_win, reflect_back, commitment_confirm |
| segments[].moment | string | Key moment, if any | e.g. insight, breakthrough, resistance |
| segments[].overlap | boolean | Whether speakers talked over each other | |
| segments[].acoustics | object | Words per minute, average pitch, pitch variation, relative loudness, pause before | Can be empty for very short segments |
| emotion_notes[] | array | Emotion shifts: speaker, from tone, to tone, the segment that triggered it, and a note | |
| transitions[] | array | Turning points in the conversation, with a note | e.g. coach_redirect, topic_shift, breakthrough, student_pushback |
| emotional_arcs[] | array | Each speaker's emotional path through the call, point by point | Score from -3 to +3 |
| behavior_effects[] | array | How each technique landed: the student, their emotional score before and after, and the direction | Direction: up, flat, down, no_response |
| redaction_summary | object | Count of each kind of detail replaced | e.g. person, business, product, location, date, amount, contact |
| qc | object | Privacy and consistency checks run on the call, and their results | |
| human_checked | boolean | Whether a person has reviewed the call | Currently false for all calls |
| consent_status, consent_terms_version | string | Consent record for the call | |
| schema_version, taxonomy_version, pipeline_version | string | Versions of the file format, label set and pipeline | |