koalareads/kr-vocab-synth-20260327-v4

KoalaReads German Vocabulary Trainer - Synthetic Conversations

Overview

This dataset contains high-quality synthetic conversational data designed for training AI-powered vocabulary trainers for German language learning. The conversations are carefully structured to follow proven pedagogical principles and the CEFR (Common European Framework of Reference) levels A1-C2, making them ideal for fine-tuning language models that need to teach and assess vocabulary acquisition.

Key Features

• Large-scale: 5,000 training conversations with natural error patterns
• CEFR-aligned: Full coverage from beginner (A1) to mastery (C2) levels
• Diverse exercises: 9 distinct exercise types targeting different learning modalities
• Pedagogically sound: Immediate feedback, error correction, and meaning reinforcement
• Production-ready: Train/validation/test splits, comprehensive metadata, validated format
• OpenAI-compatible: Chat format ready for fine-tuning with GPT, Llama, Mistral, etc.

Dataset Lineage

This dataset is generated from the official KoalaReads vocabulary:

• Source: koalareads/german-vocabulary-cefr-20260327
• Words: 3,092 German vocabulary entries with rich linguistic metadata
• Quality: CEFR-classified, manually curated, multi-source validated

Dataset Statistics

Core Metrics

Metric	Value
Total Records	5,000
Total Messages	20,822
Vocabulary Words	3,092
Avg Messages/Record	4.16
Train Split	4,000 (80%)
Validation Split	500 (10%)
Test Split	500 (10%)
Lexical Error Rate	25.3%
Generation Date	2026-04-04

Quality Indicators

• ✅ 100% validated: All records passed structural and pedagogical validation
• ✅ Realistic errors: 1,265 simulated learner mistakes for robust training
• ✅ Grammar-aware: 638 grammar errors correctly identified (not penalized as lexical)
• ✅ Immediate feedback: Every learner response followed by instructor correction
• ✅ Noise-free: All timing markers, pauses, and script directions removed

LLM-as-a-Judge Quality Gate

Every record was evaluated by gpt-4o-mini on 5 quality criteria. Records scoring below 3.5/5.0 were rejected.

Metric	Value
Records Evaluated	6,239
Records Passed	5,071
Records Rejected	1,168
Pass Rate	81.3%
Judge Model	gpt-4o-mini
Score Threshold	3.5/5.0

Per-Criterion Average Scores (1-5 scale):

Criterion	Mean	Min	Max
Semantic Consistency	4.47	1	5
Pedagogical Correctness	4.62	1	5
Linguistic Quality	4.87	1	5
Conversation Coherence	4.88	2	5
Target Alignment	4.96	1	5

Top Rejection Reasons:

• Minor grammatical issues in the user's sentence affect clarity. (15x)
• Minor grammatical issues in the example sentence. (11x)
• Minor grammatical issues in the assistant's feedback. (9x)
• Example sentence is generic and does not demonstrate actual word usage effectively. (6x)
• The example sentence provided is generic and does not demonstrate varied usage of the target word. (6x)

CEFR Level Distribution

The dataset provides balanced coverage across proficiency levels:

CEFR Level	Records	Percentage	Description
A1	540	10.8%	Beginner - Basic phrases and common words
A2	1,513	30.3%	Elementary - Simple sentences and frequent expressions
B1	1,003	20.1%	Intermediate - Main points on familiar matters
B2	1,423	28.5%	Upper Intermediate - Complex text and abstract topics
C1	305	6.1%	Advanced - Implicit meanings and subtle expressions
C2	216	4.3%	Mastery - Effortless comprehension and expression

Exercise Type Distribution

Exercise Type	Count	Description
T1_direct_translation	1,716	English → German translation with immediate feedback
T4_reverse_translation	1,026	German → English comprehension check
MN_noun_module	820	Noun-specific with article (der/die/das) training
T3_situational_recall	561	Context-based word selection from scenarios
T5_sentence_production	412	Creative sentence generation with target word
MV_verb_module	295	Verb conjugation and usage patterns
T2_sentence_completion_block	112	Fill-in-the-blank consolidation (3 words)
review_block	53	Mixed review of 5-7 previously learned words
MA_semantic_cluster	5	Distinguish between semantically similar words

Visualizations

Exercise Types

Cefr Distribution

Pos Distribution

Sample Conversations

Below are randomly selected examples demonstrating different exercise types and CEFR levels:

Example 1: T4 Reverse Translation

CEFR Level: A2 | Target Word: Lernen | POS: noun

Click to expand conversation

{
  "messages": [
    {
      "role": "system",
      "content": "You are a vocabulary trainer. Focus on exactly ONE target word per task.\nKeep instructions short. After the learner answers, give immediate feedback:\n- say if it's correct (lexically)\n- give the correct answer (and the full correct sentence when relevant)\n- briefly reinforce the meaning/usage\nGrammar mistakes alone do not count as lexical errors.\nNever mention timing, pauses, or scheduling."
    },
    {
      "role": "assistant",
      "content": "What does the German word 'Lernen' mean in English?"
    },
    {
      "role": "user",
      "content": "It means 'learning'."
    },
    {
      "role": "assistant",
      "content": "Correct! 'Lernen' translates to 'learning', which is the process through which knowledge or skills are acquired."
    }
  ],
  "metadata": {
    "schema_version": "kr_vocab_synth_v1",
    "cefr": "A2",
    "trainer": "A1-A2",
    "mode": "audio",
    "stage": 2,
    "exercise_type": "T4_reverse_translation",
    "target_word": "Lernen",
    "pos": "noun",
    "lexical_error_count": 0,
    "grammar_error_present": false,
    "meaning_demonstrated": true,
    "noise_stripped": true
  }
}

Example 2: T3 Situational Recall

CEFR Level: A2 | Target Word: sogar | POS: adv

Click to expand conversation

{
  "messages": [
    {
      "role": "system",
      "content": "You are a vocabulary trainer. Focus on exactly ONE target word per task.\nKeep instructions short. After the learner answers, give immediate feedback:\n- say if it's correct (lexically)\n- give the correct answer (and the full correct sentence when relevant)\n- briefly reinforce the meaning/usage\nGrammar mistakes alone do not count as lexical errors.\nNever mention timing, pauses, or scheduling."
    },
    {
      "role": "assistant",
      "content": "Imagine you're at a birthday party, and you see a huge cake with multiple layers. Your friend says they made a cake that is not just big but even has extra surprises inside. What word fits best in German?"
    },
    {
      "role": "user",
      "content": "sogar"
    },
    {
      "role": "assistant",
      "content": "That's right! \"Sogar\" means \"even,\" and it emphasizes the surprise about the cake. For example, you could say: \"Der Kuchen hat sogar Überraschungen im Inneren.\""
    }
  ],
  "metadata": {
    "schema_version": "kr_vocab_synth_v1",
    "cefr": "A2",
    "trainer": "A1-A2",
    "mode": "audio",
    "stage": 2,
    "exercise_type": "T3_situational_recall",
    "target_word": "sogar",
    "pos": "adv",
    "lexical_error_count": 0,
    "grammar_error_present": false,
    "meaning_demonstrated": true,
    "noise_stripped": true
  }
}

Example 3: T4 Reverse Translation

CEFR Level: B2 | Target Word: Angetrieben | POS: verb

Click to expand conversation

{
  "messages": [
    {
      "role": "system",
      "content": "You are a vocabulary trainer. Focus on exactly ONE target word per task.\nKeep instructions short. After the learner answers, give immediate feedback:\n- say if it's correct (lexically)\n- give the correct answer (and the full correct sentence when relevant)\n- briefly reinforce the meaning/usage\nGrammar mistakes alone do not count as lexical errors.\nNever mention timing, pauses, or scheduling."
    },
    {
      "role": "assistant",
      "content": "What does the German word 'Angetrieben' mean in English?"
    },
    {
      "role": "user",
      "content": "It mean driven."
    },
    {
      "role": "assistant",
      "content": "That's right! 'Angetrieben' translates to 'driven'. Just remember to use correct grammar; it should be 'It means driven.' Great job with the vocabulary!"
    }
  ],
  "metadata": {
    "schema_version": "kr_vocab_synth_v1",
    "cefr": "B2",
    "trainer": "B1-C2",
    "mode": "digital",
    "stage": 2,
    "exercise_type": "T4_reverse_translation",
    "target_word": "Angetrieben",
    "pos": "verb",
    "lexical_error_count": 0,
    "grammar_error_present": true,
    "meaning_demonstrated": true,
    "noise_stripped": true
  }
}

Example 4: T1 Direct Translation

CEFR Level: A2 | Target Word: Gefühl | POS: noun

Click to expand conversation

{
  "messages": [
    {
      "role": "system",
      "content": "You are a vocabulary trainer. Focus on exactly ONE target word per task.\nKeep instructions short. After the learner answers, give immediate feedback:\n- say if it's correct (lexically)\n- give the correct answer (and the full correct sentence when relevant)\n- briefly reinforce the meaning/usage\nGrammar mistakes alone do not count as lexical errors.\nNever mention timing, pauses, or scheduling."
    },
    {
      "role": "assistant",
      "content": "How do you say 'feeling' in German?"
    },
    {
      "role": "user",
      "content": "Das Gefühl."
    },
    {
      "role": "assistant",
      "content": "Great job! You got the word right: 'das Gefühl' means 'the feeling.' Just remember to use it in the right context. For example: 'Ich habe ein komisches Gefühl,' which means 'I have a strange feeling.'"
    }
  ],
  "metadata": {
    "schema_version": "kr_vocab_synth_v1",
    "cefr": "A2",
    "trainer": "A1-A2",
    "mode": "audio",
    "stage": 1,
    "exercise_type": "T1_direct_translation",
    "target_word": "Gefühl",
    "pos": "noun",
    "lexical_error_count": 0,
    "grammar_error_present": true,
    "meaning_demonstrated": true,
    "noise_stripped": true
  }
}

Example 5: Mn Noun Module

CEFR Level: B1 | Target Word: Armut | POS: noun

Click to expand conversation

{
  "messages": [
    {
      "role": "system",
      "content": "You are a vocabulary trainer. Focus on exactly ONE target word per task.\nKeep instructions short. After the learner answers, give immediate feedback:\n- say if it's correct (lexically)\n- give the correct answer (and the full correct sentence when relevant)\n- briefly reinforce the meaning/usage\nGrammar mistakes alone do not count as lexical errors.\nNever mention timing, pauses, or scheduling."
    },
    {
      "role": "assistant",
      "content": "What is the German word for 'poverty'?"
    },
    {
      "role": "user",
      "content": "I don't know."
    },
    {
      "role": "assistant",
      "content": "The correct word is 'die Armut'. It remains the same in plural. For example, you can use it in sentences like, 'Die Armut in der Stadt ist ein großes Problem' (The poverty in the city is a big issue) or 'Viele Menschen leben in Armut' (Many people live in poverty)."
    }
  ],
  "metadata": {
    "schema_version": "kr_vocab_synth_v1",
    "cefr": "B1",
    "trainer": "B1-C2",
    "mode": "digital",
    "stage": 1,
    "exercise_type": "MN_noun_module",
    "target_word": "Armut",
    "pos": "noun",
    "lexical_error_count": 1,
    "grammar_error_present": false,
    "meaning_demonstrated": false,
    "noise_stripped": true
  }
}

Pedagogical Design

Core Principles

This dataset implements evidence-based language teaching methodologies:

1. Immediate Feedback Loop

Every learner response receives instant, structured feedback:

• ✅ Correctness assessment (lexical accuracy prioritized)
• 📝 Correct answer provided explicitly
• 🔄 Full correct sentence given for sentence-based exercises
• 💡 Meaning reinforcement with examples

2. Progressive Difficulty

Exercises are sequenced by CEFR level:

• A1-A2: Simple translation, basic recall, common vocabulary
• B1-B2: Situational usage, sentence production, nuanced meanings
• C1-C2: Semantic distinctions, implicit meanings, complex structures

3. Multi-Modal Learning

Nine exercise types target different cognitive processes:

• Recognition (T4: German → English)
• Recall (T1: English → German)
• Application (T3: Situational contexts)
• Production (T5: Sentence creation)
• Integration (T2: Consolidation blocks)

4. Error-Aware Training

Realistic learner mistakes are simulated:

• Lexical errors (wrong word choice)
• Grammar errors (correctly not penalized for vocabulary tasks)
• Typos and spelling mistakes
• "I don't know" responses
• Ambiguous or off-topic answers

Error Simulation Distribution

The dataset includes realistic learner responses:

Response Type	Approximate Rate	Purpose
Correct	~77%	Reinforce positive patterns
Lexical Error	~15%	Teach correction strategies
Grammar Error	~12%	Train grammar tolerance
Typo	~8%	Handle spelling variations
Don't Know	~5%	Model uncertainty handling
Off-Topic	~2%	Teach redirection

✨ Note: Rates overlap; responses can have multiple error types.

Exercise Types - Detailed Reference

T1: Direct Translation (English → German)

Purpose: Basic word recall and production
Format: Simple translation prompt with immediate feedback

Assistant: How do you say "to go" in German?
User: gehen
Assistant: ✅ Correct: "gehen".
  Meaning: "to go".
  Example: Ich gehe nach Hause.

T2: Sentence Completion Block

Purpose: Consolidation after learning 3 words
Format: Fill-in-the-blank with word bank

Assistant: Fill the blanks using: gehen, Zeit, Wasser
  1) Ich ____ nach Hause.
  2) Die ____ ist wichtig.
  3) Das ____ ist kalt.
User: 1) gehen 2) Zeit 3) Wasser
Assistant: [Provides complete sentences as confirmation]

T3: Situational Recall

Purpose: Context-based word activation
Format: Realistic scenario → appropriate word

Assistant: Situation: You are leaving work and going home on foot.
  What word fits best in German?
User: gehen
Assistant: ✅ Yes: "gehen".
  Example sentence: Ich gehe jeden Tag zu Fuß.

T4: Reverse Translation (German → English)

Purpose: Comprehension check
Format: Recognition and meaning recall

Assistant: What does "gehen" mean in English?
User: to go
Assistant: ✅ Correct! "gehen" = "to go".

T5: Sentence Production

Purpose: Active usage in context
Format: Creative sentence with target word

Assistant: Make a simple sentence using "gehen".
  Example format: [Subject] [verb] [complement]
User: Ich gehe ins Kino.
Assistant: ✅ Excellent! "Ich gehe ins Kino." uses "gehen" correctly.

MN: Noun Module

Purpose: Article (der/die/das) + noun mastery
Format: Noun with correct article

MV: Verb Module

Purpose: Conjugation and verb patterns
Format: Verb with principal parts (infinitive, past, perfect)

MA: Semantic Cluster

Purpose: Distinguish similar words
Format: Multiple situations requiring different near-synonyms

Review Block

Purpose: Spaced repetition of learned words
Format: 5-7 mixed exercises from previous lessons

Dataset Structure

Data Format

Each record is a complete conversation in OpenAI chat format, ready for fine-tuning:

{
  "messages": [
    {
      "role": "system",
      "content": "You are a vocabulary trainer..."
    },
    {
      "role": "assistant",
      "content": "How do you say \"water\" in German?"
    },
    {
      "role": "user",
      "content": "Wasser"
    },
    {
      "role": "assistant",
      "content": "✅ Correct: \"Wasser\". Meaning: \"water\". Example: Das Wasser ist kalt."
    }
  ],
  "metadata": {
    "schema_version": "kr_vocab_synth_v1",
    "cefr": "A1",
    "exercise_type": "T1_direct_translation",
    "target_word": "Wasser",
    "target_word_id": "word_12345",
    "pos": "noun",
    "lexical_error_count": 0,
    "grammar_error_present": false,
    "meaning_demonstrated": true
    // ... additional metadata ...
  }
}

Metadata Fields

Rich metadata enables advanced filtering and analysis:

Core Identification

• schema_version: Dataset version (kr_vocab_synth_v1)
• exercise_type: One of 9 exercise types
• target_word: German word being trained (lemma form)
• target_word_id: Unique vocabulary ID (links to vocab dataset)
• source_text_id: Reading/text where word appears (if applicable)

CEFR & Progression

• cefr: Target CEFR level (A1, A2, B1, B2, C1, C2)
• trainer: Trainer type ("A1-A2" or "B1-C2")
• stage: Progressive stage number (1-4)
• mode: Training mode (typically "audio")

Linguistic Information

• pos: Part of speech (noun, verb, adj, adv, etc.)
• article_de: German article (der, die, das) for nouns
• plural_de: Plural form for nouns

Quality Metrics

• lexical_error_count: Number of vocabulary mistakes in response
• grammar_error_present: Boolean for grammar mistakes
• meaning_demonstrated: Whether meaning was shown
• noise_stripped: Confirmation that timing noise removed

File Organization

dataset/
├── train/              # 90% of data (18,000 records for 20K dataset)
│   ├── dataset_info.json
│   ├── state.json
│   └── data-*.arrow
└── test/               # 10% of data (2,000 records for 20K dataset)
    ├── dataset_info.json
    ├── state.json
    └── data-*.arrow

Usage Guide

Quick Start

from datasets import load_dataset

# Load full dataset
dataset = load_dataset('koalareads/kr-vocab-synth-20260327-v1')

# Access splits
train_data = dataset['train']
test_data = dataset['test']

# Inspect a record
print(train_data[0])
# {'messages': [...], 'metadata': {...}}

Fine-Tuning Preparation

For OpenAI API

import json

# Convert to JSONL for OpenAI fine-tuning
with open('training_data.jsonl', 'w') as f:
    for record in dataset['train']:
        f.write(json.dumps(record) + '\n')

For Hugging Face Transformers

from transformers import AutoTokenizer, AutoModelForCausalLM

model_name = "mistralai/Mistral-7B-v0.1"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForCausalLM.from_pretrained(model_name)

# Apply chat template
def format_messages(example):
    return {
        "text": tokenizer.apply_chat_template(
            example['messages'], 
            tokenize=False
        )
    }

formatted_dataset = dataset.map(format_messages)

Filtering and Analysis

# Filter by CEFR level
a1_data = dataset['train'].filter(lambda x: x['metadata']['cefr'] == 'A1')

# Filter by exercise type
translation_data = dataset['train'].filter(
    lambda x: x['metadata']['exercise_type'] == 'T1_direct_translation'
)

# Filter by part of speech
noun_exercises = dataset['train'].filter(
    lambda x: x['metadata']['pos'] == 'noun'
)

# Get error examples for robust training
error_examples = dataset['train'].filter(
    lambda x: x['metadata']['lexical_error_count'] > 0
)

print(f"A1 records: {len(a1_data)}")
print(f"Translation exercises: {len(translation_data)}")
print(f"Noun exercises: {len(noun_exercises)}")
print(f"Error examples: {len(error_examples)}")

Integration Examples

1. Build a Vocabulary Trainer API

# Fine-tune a model on this dataset, then:
from openai import OpenAI
client = OpenAI()

response = client.chat.completions.create(
    model="ft:gpt-3.5-turbo:koalareads::vocab-trainer",
    messages=[
        {"role": "system", "content": "You are a vocabulary trainer..."},
        {"role": "assistant", "content": "How do you say 'house' in German?"},
        {"role": "user", "content": "Haus"}
    ]
)
print(response.choices[0].message.content)
# ✅ Correct: "Haus". Meaning: "house". Example: Das Haus ist groß.

2. Create Adaptive Learning Paths

def get_exercises_for_level(cefr_level, n=10):
    """Get diverse exercises for a specific CEFR level"""
    level_data = dataset['train'].filter(
        lambda x: x['metadata']['cefr'] == cefr_level
    )
    return level_data.shuffle().select(range(n))

# Generate A1 practice set
beginner_exercises = get_exercises_for_level('A1', n=20)

3. Build Assessment Tools

# Use validation split for model selection, test split for final evaluation
test_records = dataset['test']

# Evaluate model performance by CEFR level
for level in ['A1', 'A2', 'B1', 'B2', 'C1', 'C2']:
    level_test = [r for r in test_records if r['metadata']['cefr'] == level]
    print(f"{level}: {len(level_test)} test cases")

Use Cases

1. Fine-Tuning Language Models

Train small, efficient models (Mistral-7B, Llama-3-8B, GPT-3.5) to act as vocabulary trainers:

• Immediate feedback generation
• Error correction strategies
• CEFR-appropriate language
• Pedagogical conversation flow

2. Conversational AI for Education

Build interactive learning assistants:

• Vocabulary quiz bots
• Adaptive tutoring systems
• Spaced repetition schedulers
• Progress tracking dashboards

3. Research & Development

• Study error correction patterns
• Analyze pedagogical effectiveness
• Benchmark model performance on educational tasks
• Experiment with different training strategies

4. Dataset Augmentation

• Combine with real learner conversations
• Bootstrap cold-start systems
• Generate additional training data with variations

📈 Expected Model Performance

Models fine-tuned on this dataset typically learn to:

✅ Generate immediate, structured feedback (high precision)
✅ Distinguish lexical vs. grammar errors (critical for vocab training)
✅ Provide correct answers with examples (pedagogically sound)
✅ Adapt tone and complexity to CEFR level (A1 simpler than C2)
✅ Handle "I don't know" and off-topic responses (robust)

Recommended Training Setup

• Model size: 7B-13B parameters (Mistral, Llama, GPT-3.5)
• Training epochs: 2-3 (avoid overfitting)
• Learning rate: 1e-5 to 5e-5
• Batch size: 4-8 (depending on GPU)
• Data splits: Train (80%), validation (10%), test (10%)
• Evaluation: Use validation split for model selection, test split for final metrics

Citation

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

@dataset{koalareads_vocab_synth_2026,
  title = {KoalaReads German Vocabulary Trainer - Synthetic Conversations},
  author = {KoalaReads Team},
  year = {2026},
  month = {4},
  publisher = {HuggingFace},
  howpublished = {\url{https://huggingface.co/datasets/koalareads/kr-vocab-synth-20260327-v1}},
  note = {Version 1, 5,000 synthetic conversations based on CEFR-classified German vocabulary}
}

Related Work

Please also cite the source vocabulary dataset:

@dataset{koalareads_german_vocab_2026,
  title = {KoalaReads German Vocabulary with CEFR Classifications},
  author = {KoalaReads Team},
  year = {2026},
  publisher = {HuggingFace},
  howpublished = {\url{https://huggingface.co/datasets/koalareads/german-vocabulary-cefr-20260327}}
}

License

MIT License

This dataset is released under the MIT License, making it free to use for:

• ✅ Educational purposes
• ✅ Research and academic work
• ✅ Commercial applications
• ✅ Derivative works and modifications

Terms

• Attribution: Please cite this dataset and the source vocabulary
• No Warranty: Provided "as is" without guarantees
• Freedom to Use: Modify, distribute, and commercialize freely

See LICENSE file for full terms.

Contributing & Feedback

We welcome contributions and feedback!

Report Issues

• Bugs: Found data quality issues? Open an issue
• Suggestions: Ideas for improvement? Share them on our discussions page
• Questions: Need help using the dataset? Ask in the community tab

Contribute

• Extend: Generate additional exercise types
• Improve: Enhance the model card or examples
• Validate: Test with your models and share results

Contact

• Website: koalareads.com
• GitHub: koalareads/koalareads-datasets
• Email: datasets@koalareads.com

Version History

v1 (Current)

• 5,000 synthetic conversations
• 3,092 vocabulary words from koalareads/german-vocabulary-cefr-20260327
• 9 exercise types with pedagogical validation
• Full CEFR coverage (A1-C2)
• Realistic error simulation

Future Versions

Planned improvements for v2+:

• Multi-turn extended conversations
• Pronunciation feedback exercises
• Cultural context integration
• Adaptive difficulty modeling
• Real learner feedback integration

Acknowledgments

This dataset was created by the KoalaReads team with contributions from:

• Linguistic experts for CEFR classification
• Language teachers for pedagogical validation
• Open-source community for tools and infrastructure

Built with:

• 🤗 Hugging Face Datasets
• 🐍 Python 3.12+
• 📊 Matplotlib for visualizations

---

Generated: 2026-04-04
Quality: Production-ready, validated
Status: Actively maintained

Happy training! 🚀🎓