🧸 EleMo-V1 (Elementary Pedagogical Model GERMAN)
Status: V-1 Release Version (Production Ready)
Model Description
EleMo (Elementarpädagogisches Modell) is the first highly specialized, locally deployable Large Language Model (LLM) tailored exactly to the requirements of early childhood education.
This model is designed to solve a core dilemma: pedagogical professionals are drowning in documentation duties, while lacking time for the children. However, conventional, cloud-based AI models are completely out of the question in this sensitive environment. We must not upload the most intimate developmental data of young children to external servers.
EleMo provides the alternative: Local, data-sovereign AI for elementary pedagogy.
Intended Use & Guardrails
A general language model does not know how to document professionally. It tends to make judgments, fabrications, and inappropriate diagnostics. EleMo is strictly fine-tuned to process raw, bullet-pointed observation data into authentic Learning Stories (Lerngeschichten).
In doing so, the model follows strict pedagogical guardrails anchored in the specific observation model by Margaret Carr:
- Methodology of Learning Stories (Margaret Carr): EleMo is deeply aligned with Carr's observation and documentation framework. The narrative output is consistently structured according to the triad: Observation – Significance – Next Steps (Opportunities), ensuring that the child's learning dispositions are made visible without deficit-oriented assessment.
- Factual Accuracy: Every statement about learning must be based on a previously described observation. What has not been observed must not be claimed or hallucinated.
- The Right Tonality: The Learning Story is not a dry report, not a cold analysis, and certainly no developmental diagnostics. It is written as a personal, appreciative letter directly to the child.
📊 Dataset Balance & Bias Mitigation Matrix (v1)
To ensure high pedagogical quality and prevent systemic biases (such as gender stereotypes or a pure positivity bias) right from the start, the foundational training dataset for EleMo-v1 was mathematically balanced across multiple dimensions.
1. Dataset Distribution Matrix
The v1 dataset consists of 20 meticulously curated pedagogical scenarios distributed across the following core dimensions:
| Dimension | Category / Focus | Count | Share (%) |
|---|
| Educational Areas | Contribution, Exploration, Identity, Communication, Creativity, Physical Development, Mathematical Thinking, Social Relationships, Environmental Awareness, Well-being | 2 each | 10% each |
| Age Groups | 2–3 Years (20.5%), 3–4 Years (25%), 4–5 Years (25%), 5–6 Years (29.5%) | 20 total | 100% |
| Gender Balance | Boys (34.5%), Girls (34%), Diverse (31.5%) | 20 total | 100% |
| Location | Outdoors (45.5%), Indoors (53%), Circle Time (1.5%) | 20 total | 100% |
| Family Structures | Nuclear Family (62%), Single Parent (15%), Extended Family (9.5%), Patchwork (9.5%), Same-Sex Parents (4%) | 20 total | 100% |
Bias Mitigation Strategy
We actively engineered the dataset to solve common LLM alignment issues in early childhood documentation:
- Gender & Role Stereotypes: Crossed all genders thoroughly with all educational areas (e.g., breaking the pattern of boys exclusively in math/exploration and girls in creativity). Gender distribution is leveled equally (~33% each).
- Cultural & Naming Diversity: Explicitly differentiated subcultures (e.g., specific Arabic cultural backgrounds like Egypt, Morocco, Lebanon, Syria) and decoupled the "Cultural Background" column from implicit naming assumptions to prevent token-level stereotyping.
- Emotional Realism (Countering Positivity Bias): Standard models struggle to document negative emotions pedagogically. We intentionally expanded scenarios involving Frustration, Disappointment, Sadness, and Loneliness to train the model on how to handle difficult emotional situations within the Margaret Carr framework without assessing them negatively.
- Inclusion Markers: Integrated explicitly defined inclusion features (AAC communication aids, motor/visual/hearing impairments, chronic illnesses) to ensure the model naturally represents inclusive environments.
The Zero-Cloud Guarantee
We rely on the principle of data sovereignty through local execution. This model is built to run directly on-site, on local hardware, or in strictly isolated, sovereign environments. What happens in the kindergarten stays in the kindergarten.
Scientific Access
Since EleMo embodies deep-seated methodical knowledge of elementary pedagogy, it is currently maintained as a Private / Gated Model. To prevent these specialized weights from flowing unchecked into commercial software products, access is strictly regulated.
— Sebastian Götz | Owner of Kita Digital & Initiator of KI-Insel
