SakhrML/SpeakMK1_SLP_Dialogue

Dataset Card for SLP Dialogue Dataset

Dataset Description

This dataset consists of 1,000 multi-turn simulated pediatric Speech-Language Pathology (SLP) interaction dialogues. It is specifically designed to train, evaluate, or fine-tune LLMs to act as clinical speech-language therapists or to study clinical reasoning during speech therapy sessions. Each conversation includes clinical metadata (child age, speech sound disorder category, specific phone error, clinical goal, complexity level, and therapeutic strategy), explicit sequential reasoning paths (model_thought), and the target therapeutic response (slp_response).

• Total Dialogues: 1,000
• Total Conversational Turns: 3,625
• Average Turns per Dialogue: 3.62
• Target Population: Children aged 3–10 presenting with various speech sound disorders.

---

Dataset Structure

Data Instances

Each row in the dataset is a JSON object containing the metadata configuration, an array of dialogue turns representing the interaction, and a generation seed.

{
  "metadata": {
    "age": 5,
    "disorder": "Interdental Lisp",
    "specific_error": "Frontal interdental /s/ (s → θ)",
    "clinical_goal": "Build dental occlusion awareness for sibilants",
    "primary_strategy": "Minimal Pair Contrast",
    "complexity_level": 1
  },
  "dialogue": [
    {
      "turn": 1,
      "child_input": "Look! I got a new thun! It's thuper thiny and thparkly!",
      "model_thought": "Child is exhibiting interdental /s/ substitution (s → θ) in 'sun', 'super', and 'sparkly'. The child is highly motivated by a new object (sun/sticker). Strategy: Use minimal pair contrast to highlight the difference between the 'leaky' /θ/ and the 'tight' /s/...",
      "slp_response": "Oh my goodness, look at that! It is so, so shiny! It's like a little piece of a star! [gasps] But wait, I heard a little bit of air escaping—it sounded a bit like 'thun.' Let's try a magic trick with our teeth..."
    }
  ],
  "_seed": 42
}

Data Fields

metadata

• age (integer): Age of the simulated patient (ranging from 3 to 10 years old).
• disorder (string): The speech sound disorder category or phonological process being targeted (e.g., Fronting, Stopping, Interdental Lisp).
• specific_error (string): Detailed phonological mutation or substitution description (e.g., $s \rightarrow \theta$, $\int \rightarrow s$, final consonant deletion).
• clinical_goal (string): The specific therapeutic objective for the session.
• primary_strategy (string): The clinical methodology used by the SLP agent to elicit the correct speech sound.
• complexity_level (integer): Task complexity scaling from 1 (Isolation/Syllable/Single Word level) to 3 (Connected Speech/Conversation with built-in distractors).

dialogue

• turn (integer): Sequential turn index (starting at 1).
• child_input (string): Utterance from the child, modeling specific phonological errors or behavioral characteristics.
• model_thought (string): Chain-of-thought clinical rationale. Analyzes the child's errors, tracks clinical progress/regressions, and justifies the chosen intervention tactic.
• slp_response (string): The explicit therapist utterance containing positive reinforcement, age-appropriate play cues, and structural/articulatory modeling.

_seed

• _seed (integer): The random seed value used for variability during dataset creation.

---

Dataset Statistics

Disorder Distribution

The dataset covers 12 distinct speech sound disorders and phonological processes balanced evenly across the 1,000 cases:

Disorder Category	Case Count
Fronting	96
Backing	92
Stopping	88
Frontal Lisp	86
Final Consonant Deletion	85
Assimilation	85
Cluster Reduction	81
Interdental Lisp	80
Inconsistent Deviant Disorder	80
Gliding	80
Lateral Lisp	79
Palatalization	68

Age Distribution

Age	Case Count
3	83
4	157
5	248
6	192
7	156
8	84
9	57
10	23

Complexity Levels

Level	Focus Description	Case Count
1	Foundational phonetic placement & minimal contrasts	227
2	Word-initial/final target stabilization	513
3	Naturalistic natural speech with cognitive/thematic distractors	260

Primary Intervention Strategies

The conversations systematically demonstrate 12 highly utilized evidence-based clinical practices:

• Minimal Pair Contrast
• Tactile Cueing (e.g., touching alveolar ridge awareness)
• Scaffolded Expansion
• Play-Based Naturalistic Elicitation
• Phonological Awareness Activities
• Self-Monitoring Training
• Visual Demonstration (SLP models placement)
• Backward Chaining
• Contrastive Drilling with Reinforcement
• Core Vocabulary Approach

---

Intended Uses & Limitations

Applications

• Clinical Co-Pilot Training: Training models to assist speech therapists in structuring fun, conversational interventions.
• Reasoning-Focused LLMs: Benchmark or train models to perform implicit multi-step clinical tracking (model_thought) prior to conversational output formulation.
• Synthetic Patient Simulation: Evaluating an AI clinician's ability to remain patient, encouraging, and target-focused when confronted with child-like digressions.

Limitations

• This dataset contains text representations of speech errors (e.g., writing "thun" for "sun"). Real-world speech therapy relies heavily on audio acoustics and visual observation which text transcriptions cannot fully capture.
• The data is synthetically generated and should be validated by certified clinical professionals prior to integration in high-stakes clinical deployments.

---

How to Load

Using the Hugging Face datasets library

Once uploaded to the Hugging Face Hub, users can load the dataset directly via:

from datasets import load_dataset

dataset = load_dataset("your-username/slp-dialogue-stage3")
print(dataset['train'][0])

Loading the Local JSONL File

import json

with open("slp_stage3_train_clean.jsonl", "r", encoding="utf-8") as f:
    for line in f:
        data_instance = json.loads(line)
        # Process data_instance