🏆 EleMo-V3 (Elementary Pedagogical Model)
Choose your language / Wähle deine Sprache: - 🇩🇪 Deutsch - 🇬🇧 English - 🇬🇧 English-US
🇺🇸 US English
Status: V-3 Advanced Production Version (N=500 Balanced Dataset / N=2000+ Platinum-Standard)
Model Description EleMo-V3 represents the evolutionary milestone of the Elementary Pedagogical Model. It is a highly specialized, locally executable Large Language Model (LLM) tailored exactly to the requirements of data-sovereign early childhood documentation.
A comparative deep dive shows that EleMo V3, through its specialization, dominates in a child-friendly, empathetic tone and a focus on social co-construction. While generalist frontier models exhibit high analytical sharpness, they often fall back into sterile, reporting adult language. EleMo V3, on the other hand, writes learning stories based on Margaret Carr's framework as authentic, appreciative letters to the child, making their learning paths and self-efficacy visible in a resource-oriented way to build a strong pedagogical relationship.
Furthermore, EleMo offers a strict "Zero-Cloud Guarantee." The model runs completely locally to process sensitive child development data in compliance with GDPR and without a cloud connection. Detailed insights can be found in the document Deep_Dive_Modellvergleich_Lerngeschichten.pdf.
Please note: The free EleMo variants are not as powerful as the commercial versions due to lower quantization!
🔐 Licensing & Model Variants EleMo-V3 is available in various model and quantization variants. Licensing terms depend on the respective variant.
🟢 Free Community Variants — Non-Commercial The following quantized variants are available as free downloads for private and non-commercial use:
- Q5, Q4, Q3, Q2, and other quantizations below Q6
These versions are intended for:
- Private use & non-commercial educational purposes
- Research, experiments & community projects
- Local deployment and testing Commercial use is not permitted under the free license.
🔒 Licensed Variants — Q6 & Q8 The higher-quality Q6 & Q8 variants are paid, licensed versions of EleMo-V3.
- They are not released under the free Non-Commercial license.
- A separate license is strictly required for: Commercial use, commercial deployment, integration into commercial products/services, commercial redistribution, or hosting/provisioning as a service. Important: Downloading or receiving a Q8 or BF16 model file does not grant commercial usage rights.
📊 Scientific Framework & Dataset Distribution Matrix To eliminate systemic algorithmic biases and reflect complex pedagogical realities, the training dataset was constructed across 8 specific scientific dimensions:
- Core Educational Areas (Strict Equal Distribution): The 10 educational areas (e.g., well-being, exploration, communication, social relationships) are mathematically weighted equally (10% each) to prevent topic-specific bias.
- Demographics (Age & Balanced Gender Rotation): Age groups from 2–6 years are evenly distributed. Gender balance: Girls (33.4%), Boys (33.4%), Diverse (31.2%). Maximally balanced to break down role stereotypes.
- Cultural & Linguistic Diversity: Explicit separation of cultural spheres to avoid token-level naming bias (e.g., German, Turkish, Arabic/Syria, Arabic/Morocco, Polish, Vietnamese).
- Modern Family Structures: Dismantling the unreflected bias of the nuclear family as the sole standard (Nuclear Family 50%, Single Parent 20%, Blended Family 15%, Multi-Generational 10%, Same-Sex Parents 5%).
- Inclusion, Neurodivergence & Language Acquisition: Anchored in inclusive pedagogy (including German as a Second Language 12%, Language Delay 8%, High Sensitivity 8%, Suspected ADHD 6%, Autism Spectrum 6%).
- Locations & Micro-Transitions (Transition Research): Capturing critical pedagogical everyday transitions according to Griebel & Niesel (morning circle, cloakroom, meal situations, diaper changing).
- Emotional Range (Mitigating Positivity Bias): Systematic integration of challenging emotions to cope with crisis moments supported by resilience research (anger, frustration, sadness).
- Learning Dispositions (Margaret Carr Framework): Theoretical foundation for educational and learning stories (being interested, being involved, persisting with difficulty, expressing ideas, taking responsibility).
🛠️ Dataset Engineering & Bias Mitigation Matrix As documented in the scientific justification repository (Ref: EleMo_Goldstandard_3000_Matrix_v3.xlsx):
| Category | Original LLM Problem | Revised Solution (V3) | Scientific Framework |
|---|---|---|---|
| Emotional Range | Strong focus on positive emotions (pride, joy). Complex feelings are missing. | Systematic integration of defiance, anger, sadness, and frustration. | Resilience Research: Coping with frustration as a core competency (Wustmann). |
| Inclusion Markers | 96.5% baseline without neurodivergence or language barriers. | Structural representation of German as a Second Language, Language Development Delays, high sensitivity, ADHD, and autism. | Inclusive Pedagogy: Depicting authentic heterogeneity in everyday daycare life. |
| Locations & Transitions | Almost exclusively binary separation (Indoors/Outdoors). | Comprehensive mapping of the cloakroom, changing table, dining table, and pick-up scenarios. | Transition Research: Micro-transitions as critical learning opportunities (Griebel/Niesel). |
| Pedagogical Structure | Simple metadata grid without theoretical depth. | Structured pillars: Learning Dispositions (Carr), Objective Observation, Meaning, Direct Quotes. | Margaret Carr (2001): Assessment in Early Childhood Settings (Learning Stories). |
Example training dataset at the end of the Readme!
🛑 Intended Use & Guardrails EleMo-V3 converts bulleted, objective observation data into highly appreciative, personal letters directly to the child. It reduces strict clinical diagnostics, deficit-oriented text generation, or uncontrolled hallucinations. Every statement from the model should be explicitly anchored in the user's objective observations.
🎓 Scientific Access EleMo-V3 is maintained as a Private / Gated Model to protect the integrity of the underlying dataset matrix and prevent commercial exploitation. Non-commercial, academic research access can be requested via the institutional request forms.
— Sebastian Götz | Owner of Kita Digital & Initiator of KI-Insel
📝 Example Training Data Point (System, User, Assistant)
{
"messages": [
{
"role": "system",
"content": "Create an educational and learning story based on the provided raw data. The story is based on Margaret Carr's methodology and consistently follows the principle: 1. Observation - 2. Meaning - 3. Future Opportunities. It is not a report, an analysis, or a developmental diagnosis, but a personal letter to the child. Every statement about learning must be based on a previously described observation [...]"
},
{
"role": "user",
"content": "Filip is sitting cross-legged on a large cushion in the role-play area. In front of him is a box with dolls and clothes. [...] He takes a rag doll and a yellow dress with buttons out of the box. He places the doll on the floor in front of him and pulls the dress over its head. The doll's right arm gets caught in the sleeve. Filip feels with his fingers from the outside through the fabric of the dress, grabs the doll's hand, and pushes it outside through the sleeve opening. [...] He shouts in a loud voice: 'Gotowe! My baby is warm!'."
},
{
"role": "assistant",
"content": "Dear Filip, this morning you were sitting cross-legged on a large cushion in the role-play area. [...] Then you took a rag doll and a yellow dress with buttons out of the box. [...] When the doll's arm got caught in the sleeve and you felt it from the outside through the fabric and pushed it out, it seemed to me as if you had found a targeted solution for this dressing challenge. Your upright posture and your loud exclamation suggest that finishing the dressing and holding the doll to your chest were moments of great significance for you. I invite you to join me tomorrow in the role-play area to collect more items of clothing with different fasteners, like zippers, buttons, or bows. [...]"
}
]
}