Datasetpublic

clemNova/LUCID-Socratic-10K

🧠 LUCID-Socratic-10K The First Open Dataset for Socratic AI Tutoring Aligned with the French National Curriculum A high-quality synthetic dataset of 1,000+ multi-turn Socratic dialogues, adaptive quizzes, flashcards, and revision materials across 18 subjects and 7 grade levels — purpose-built for training pedagogically-aligned AI tutors. 📊 Explore the Data · 🚀 Quick Start · 📖 Paper · 🏗️ Architecture 🎯 What Makes… See the full description on the dataset page: https://huggingface.co/datasets/clemNova/LUCID-Socratic-10K.

sourceHugging Facecc-by-nc-sa-4.0updated 4mo agoView on Hugging Face

1likes29downloads

Dataset Card

🧠 LUCID-Socratic-10K

The First Open Dataset for Socratic AI Tutoring

Aligned with the French National Curriculum

![License: CC BY-NC-SA 4.0](https://creativecommons.org/licenses/by-nc-sa/4.0/) ![Dataset Size]() ![Gold Standard]() ![Subjects]() ![Curriculum]()

A high-quality synthetic dataset of 1,000+ multi-turn Socratic dialogues, adaptive quizzes, flashcards, and revision materials across 18 subjects and 7 grade levels — purpose-built for training pedagogically-aligned AI tutors.

📊 Explore the Data · 🚀 Quick Start · 📖 Paper · 🏗️ Architecture

</div>

🎯 What Makes LUCID-Socratic-10K Unique

Most synthetic instruction-tuning datasets (Alpaca, Dolly, OpenOrca…) train models to give answers. LUCID-Socratic-10K is fundamentally different — it trains models to make students think.

Feature	Typical Instruction Datasets	LUCID-Socratic-10K
Pedagogical method	Direct Q&A	Socratic questioning — guided discovery
Curriculum alignment	Generic knowledge	French Éducation Nationale (6ème → Terminale)
Quality assurance	None or human-only	LLM-as-Judge + Human Gold labeling
Task diversity	Single-format	5 task types (tutor, quiz, flashcards, revision, homework)
Context grounding	Context-free	RAG-grounded on real course content
Language	English-centric	Native French with domain-specific terminology

Key Insight: Every tutor dialogue in this dataset is constrained by a system prompt that forbids giving direct answers. The AI must guide the student through questioning, hints, and reformulations — mimicking expert human tutoring behavior.

📊 Dataset Structure

Overview

Split	File	Entries	Description
Core	`dataset_examples.jsonl`	1,000	Multi-turn dialogues, quizzes, flashcards, revision & homework exercises
Contexts	`dataset_contexts.jsonl`	1,000	Source course materials used as RAG context for generation
Judgments	`dataset_judgments.jsonl`	1,000	LLM-as-Judge quality scores with detailed reasoning
Quizzes	`dataset_quiz_questions.jsonl`	90	Standalone MCQ questions with LaTeX solutions & distractor explanations
Sessions	`dataset_quiz_sessions.jsonl`	13	End-to-end student quiz sessions with answer tracking

Coverage

<details> <summary><b>📚 18 Subjects</b></summary>

Domain	Subjects
Sciences	Mathématiques, Physique-Chimie, SVT, Enseignement Scientifique, NSI, Sciences de l'Ingénieur, Technologie
Humanities	Français, Philosophie, Histoire-Géographie, HGGSP, HLP, SES
Languages	Anglais (LVA), Espagnol (LVB), Allemand (LVB), Italien (LVB), LLCER

</details>

<details> <summary><b>🎓 7 Grade Levels</b></summary>

Cycle	Levels
Collège (Middle School)	6ème, 5ème, 4ème, 3ème
Lycée (High School)	Seconde, Première, Terminale

</details>

<details> <summary><b>🔧 5 Task Types</b></summary>

Type	Description	Method
`tutor`	Multi-turn Socratic dialogue	Student asks → AI guides without answering
`quiz`	Multiple-choice questions	4 options + detailed solution + distractor explanations
`flashcards`	Spaced-repetition pairs	SM-2 compatible Q&A cards
`revision`	Structured revision sheets	Key concepts + exercises + self-assessment
`homework`	Practice exercises	Progressive difficulty, step-by-step solutions

</details>

🏗️ Generation Pipeline

LUCID-Socratic-10K was generated using LUCID LABS, a proprietary synthetic data forge built on a Pedagogical RAG (P-RAG) architecture:

┌─────────────────────────────────────────────────────────────────────────┐
│                        LUCID LABS — Data Forge                         │
├─────────────────────────────────────────────────────────────────────────┤
│                                                                       │
│   ┌──────────┐    ┌──────────────┐    ┌──────────────┐                │
│   │ Curriculum│───▶│   Context    │───▶│  Embedding   │                │
│   │  Source   │    │  Extraction  │    │  Vectorstore │                │
│   └──────────┘    └──────────────┘    └──────┬───────┘                │
│                                              │                        │
│                                     Semantic Retrieval                │
│                                              │                        │
│   ┌──────────┐    ┌──────────────┐    ┌──────▼───────┐                │
│   │ Socratic │◀───│   Prompt     │◀───│   P-RAG      │                │
│   │ Dialogue │    │  Engineering │    │  Contextual  │                │
│   └────┬─────┘    └──────────────┘    │   Injection  │                │
│        │                              └──────────────┘                │
│        │                                                              │
│   ┌────▼─────┐    ┌──────────────┐                                    │
│   │ LLM-as-  │───▶│  Gold Label  │──▶  dataset_examples.jsonl        │
│   │  Judge   │    │  Filtering   │                                    │
│   └──────────┘    └──────────────┘                                    │
│                                                                       │
└─────────────────────────────────────────────────────────────────────────┘

Generation Models

Stage	Model	Tokens
Content Generation	`gemini-2.5-flash`, `gemini-2.5-pro`	3,091,268
Quality Judgment	`gemini-2.5-pro` (LLM-as-Judge)	2,517,185
Total	—	5,608,453

Quality Assurance

Every generated example passes through a rigorous multi-stage quality pipeline:

1.Generation — Gemini generates content grounded in curriculum context
2.LLM-as-Judge — An independent model scores each example (1–10) with detailed written justification
3.Gold Labeling — Examples scoring ≥ 7 are marked is_gold: true
4.Result — 93.8% of examples achieve Gold status (938/1,000)

📋 Data Schema

`examples` (Core Dataset)

json

{
  "id": "uuid",
  "task_type": "tutor | quiz | flashcards | revision | homework",
  "subject": "Mathématiques",
  "level": "Seconde",
  "topic": "Vecteurs : Repérage et colinéarité",
  "context_id": "uuid → links to contexts table",
  "system_prompt": "Tu es un tuteur en {subject} pour un élève de {level}...",
  "content": {
    "context": "Course material used as RAG context",
    "conversation": [
      { "role": "user", "content": "Student question..." },
      { "role": "assistant", "content": "Socratic response..." }
    ]
  },
  "model_used": "gemini-2.5-flash",
  "tokens_used": 2833,
  "is_gold": true,
  "judge_score": 9
}

`judgments` (Quality Scores)

json

{
  "id": "uuid",
  "example_id": "uuid → links to examples",
  "score": 9,
  "is_gold": true,
  "reason": "Detailed pedagogical quality assessment in French..."
}

`quiz_questions` (Standalone MCQs)

json

{
  "question_data": {
    "topic": "Analysis",
    "subtopic": "Sequences",
    "question": "LaTeX-formatted question...",
    "options": [
      { "id": "A", "text": "...", "correct": true, "option_explanation": "..." },
      { "id": "B", "text": "...", "correct": false, "option_explanation": "Why this distractor is wrong..." }
    ],
    "solution": "Full step-by-step LaTeX solution..."
  }
}

🚀 Quick Start

Load with 🤗 Datasets

python

from datasets import load_dataset

# Load the full dataset
dataset = load_dataset("clemNova/LUCID-Socratic-10K")

# Load only Gold-standard examples
gold = dataset.filter(lambda x: x["is_gold"] == True)

# Filter by subject
math = dataset.filter(lambda x: x["subject"] == "Mathématiques")

# Filter by task type
tutoring = dataset.filter(lambda x: x["task_type"] == "tutor")

Fine-Tuning Format (ChatML)

Each tutor example can be directly converted to ChatML for fine-tuning:

python

def to_chatml(example):
    messages = [{"role": "system", "content": example["system_prompt"]}]
    for turn in example["content"]["conversation"]:
        messages.append({"role": turn["role"], "content": turn["content"]})
    return {"messages": messages}

Use with Unsloth / TRL

python

from unsloth import FastLanguageModel
from trl import SFTTrainer

# Load base model
model, tokenizer = FastLanguageModel.from_pretrained("unsloth/Qwen3-8B")

# Format dataset
def formatting_func(example):
    return tokenizer.apply_chat_template(to_chatml(example), tokenize=False)

# Train
trainer = SFTTrainer(
    model=model,
    train_dataset=gold,
    formatting_func=formatting_func,
    max_seq_length=4096,
)
trainer.train()

🔬 Intended Uses

Primary Use Cases

—Fine-tuning Socratic AI tutors that guide students through reasoning instead of giving direct answers
—Training educational chatbots aligned with the French national curriculum
—Benchmarking pedagogical quality of LLM-generated educational content
—Research on Retrieval-Augmented Generation in educational contexts

Out-of-Scope Uses

—This dataset should not be used as a replacement for human teachers
—Not suitable for medical, legal, or safety-critical educational domains
—The curriculum alignment is specific to the French Éducation Nationale system

⚠️ Limitations & Biases

—Synthetic Data — All dialogues are LLM-generated, not recorded from real student-teacher interactions
—French Curriculum Bias — Content is aligned with the French educational system and may not transfer to other national curricula without adaptation
—Model Bias — Generated using Google Gemini models, which may carry inherent biases from their training data
—Temporal Snapshot — Curriculum content reflects the 2024–2025 academic year
—Gold Threshold — The 93.8% gold rate is based on LLM-as-Judge scores ≥ 7/10; human validation was performed on a sample basis

📜 License

This dataset is released under the Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International License.

You are free to:

—Share — copy and redistribute the material
—Adapt — remix, transform, and build upon the material

Under the following terms:

—Attribution — Credit LUCID / 2Lucid
—NonCommercial — No commercial use without permission
—ShareAlike — Derivatives must use the same license

🏛️ Citation

If you use this dataset in your research, please cite:

bibtex

@misc{lucid_socratic_1k_2026,
  title        = {LUCID-Socratic-10K: A Synthetic Dataset for Training Pedagogically-Aligned AI Tutors via Socratic Retrieval-Augmented Generation},
  author       = {clemNova and 2Lucid},
  year         = {2026},
  publisher    = {Hugging Face},
  url          = {https://huggingface.co/datasets/clemNova/LUCID-Socratic-10K},
  note         = {4,157 entries across 18 subjects and 7 grade levels, 93.8\% gold-verified}
}

🔗 Links

Resource	Link
Dataset	LUCID-Socratic-10K on Hugging Face
Organization	2Lucid on GitHub
RAG Architecture Demo	Interactive 3D Visualization

Built with ❤️ by [clemNova](https://huggingface.co/clemNova) × [2Lucid](https://github.com/2Lucid)

Empowering AI to teach, not to answer.

</div>

🧠 LUCID-Socratic-10K

The First Open Dataset for Socratic AI Tutoring

Aligned with the French National Curriculum

🎯 What Makes LUCID-Socratic-10K Unique

📊 Dataset Structure

Overview

Coverage

🏗️ Generation Pipeline

Generation Models

Quality Assurance

📋 Data Schema

examples (Core Dataset)

judgments (Quality Scores)

quiz_questions (Standalone MCQs)

🚀 Quick Start

Load with 🤗 Datasets

Fine-Tuning Format (ChatML)

Use with Unsloth / TRL

🔬 Intended Uses

Primary Use Cases

Out-of-Scope Uses

⚠️ Limitations & Biases

📜 License

🏛️ Citation

🔗 Links

`examples` (Core Dataset)

`judgments` (Quality Scores)

`quiz_questions` (Standalone MCQs)