dataset_info:
features:
- name: dialog_id
dtype: string
- name: turns
list:
- name: bigram_overlap_prev
dtype: float64
- name: context_embedding
list: float64
- name: intent_label
dtype: string
- name: is_user
dtype: int64
- name: length_bucket
dtype: string
- name: nb_response_candidates
list: string
- name: readability
dtype: float64
- name: readability_score
dtype: float64
- name: role_embedding
list: int64
- name: sentiment_polarity
dtype: float64
- name: speaker
dtype: string
- name: text
dtype: string
splits:
- name: train
num_bytes: 515339977
num_examples: 13215
download_size: 458215847
dataset_size: 515339977
configs:
- config_name: default
data_files:
- split: train
path: data/train-*
Taskmaster-1 Enriched Dialog Dataset (Combined)
Overview
This dataset is a combined, enriched version of the self_dialog and woz_dialog splits from the Taskmaster-1 dataset. It consists of multi-turn, human-human and human-simulated conversations with systematic enhancements for machine learning workflows—especially dialog modeling, generation, and fine-grained evaluation.
All conversations are structured in a JSON format with consistent schema and include added semantic, linguistic, and behavioral annotations.
Enrichments Included
- Role Embedding
Each turn includes a binary role embedding:
[1, 0] for USER
[0, 1] for ASSISTANT
This makes it easier for sequence models to learn speaker turns without relying on string labels.
Use case: Improves model performance in transformer-based dialog agents by allowing role-aware generation and classification.
- Response Candidates
Each user turn is enriched with nb_response_candidates — 2 to 4 plausible assistant responses sampled from the dataset. These are not ground truth but plausible continuations.
Use case: Ideal for retrieval-based dialog training or negative sampling in response ranking tasks.
- Readability Score
Computed using Flesch-Kincaid metrics and other NLP readability formulas. Stored as readability (0–100 scale, higher = easier).
Use case: Enables analysis of language complexity and training adaptive LLMs for education, accessibility, or voice interfaces.
- Readability Grade Score
Stored as readability_score on a U.S. grade level (lower = easier to read). Especially relevant for UX tuning.
Use case: Allows controlling reading level in generation tasks or selecting user-appropriate training samples.
- Context Embedding
Each turn is augmented with a context_embedding vector (384-dim, Sentence-BERT). Represents the semantic context of the turn.
Use case: Enables plug-and-play use with FAISS-based semantic search, response reranking, and memory-augmented generation.
- Speaker Role Flags
An is_user flag is included for each turn (1 = user, 0 = assistant).
Use case: Simplifies filtering, evaluation, or role-specific metric computation.
- Utterance Length Bucketing
Each turn is labeled as:
short (<= 5 tokens)
medium (6–15 tokens)
long (> 15 tokens)
Use case: Enables sampling, curriculum learning, or model analysis across turn complexity.
- Bigram Overlap with Previous Turn
Computed as bigram_overlap_prev (float between 0 and 1). Measures lexical repetition with the preceding utterance.
Use case: Useful for:
Dialogue coherence metrics
Detecting stagnation or repetition in generated responses
Analyzing repair-based utterances
- Sentiment Polarity
Computed using a sentiment analyzer. Stored as sentiment_polarity:
Ranges from –1 (strongly negative) to +1 (strongly positive)
Use case: Enables emotion-aware generation, tone control, or training sentiment-conditioned agents.
- Format Summary
Each conversation has:
dialog_id: Unique identifier
turns: List of enriched utterances
Each turn includes:
{ "speaker": "USER", "text": "I’d like to book a table for 2", "role_embedding": [1, 0], "intent_label": "request", "nb_response_candidates": [...], "readability_score": 4.5, "context_embedding": [...], "readability": 85.6, "is_user": 1, "length_bucket": "medium", "bigram_overlap_prev": 0.2, "sentiment_polarity": 0.1 }
Suggested Use Cases
Fine-tuning LLMs for goal-oriented dialog
Training dialog state trackers and response rankers
Evaluating model outputs with context-aware metrics
Curriculum learning based on length or readability
Emotion- and intent-conditioned dialog modeling
Semantic retrieval and reranking systems
Citation
@inproceedings{48484, title = {Taskmaster-1: Toward a Realistic and Diverse Dialog Dataset}, author = {Bill Byrne and Karthik Krishnamoorthi and Chinnadhurai Sankar and Arvind Neelakantan and Daniel Duckworth and Semih Yavuz and Ben Goodrich and Amit Dubey and Kyu-Young Kim and Andy Cedilnik}, year = {2019} }