Model Card for oscarstories/lorastral24b_0604

Model Description

lorastral24b_0604 is a fine-tuned version of the Mistral Small 24B Instruct model, optimized for generating safe, engaging, and educational stories for children aged 6–12. Developed by HeyQQ GmbH for the Oscar Stories platform, LORA uses age-aligned prompts and safe-language training data to produce structured narratives suitable for primary education.

It supports German and is tailored for storytelling applications with moral, cognitive, and pedagogical goals.

Story Prompt Framework

LORA uses a structured system and user prompt strategy to ensure consistency and simplicity:

• System Prompt: Du bist ein Geschichtenerzähler für Kinder. Du schreibst kurze Geschichten mit genau 3 Absätzen. Die Geschichten sind für Kinder der Schulstufen 1-4. Der erste Absatz ist der Anfang der Geschichte, der dritte Absatz ist der Schluss. Beende die Geschichte mit 'ENDE.'

• User Prompt Template: Schreibe eine Geschichte über {article} namens {name}. {pronoun} liebt {interest}. Erkläre den Begriff {topic}. ############### Schulstufe {age_group}

Model Details

• Developed by: HeyQQ GmbH (Oscar Finance)
• License: MIT
• Model type: Transformer-based causal language model
• Base model: mistralai/Mistral-Small-24B-Instruct-2501
• Languages: German

Uses

Direct Use

• Educational storytelling for primary school (grades 1–4)

Out-of-Scope Use

• Legal, medical, or high-stakes content generation
• Unmoderated or open-ended chat applications
• Use with user-defined prompts without constraints

Here is the updated model card section for the Evaluation of the LORA model, formatted and structured for Hugging Face:

---

Evaluation

Benchmarks

The model was evaluated on standard zero-shot multiple-choice benchmarks using the Language Model Evaluation Harness (lm-eval-harness) [Gao et al., 2024]:

Task	Accuracy	± SE	Normalized Accuracy	± SE
ARC-Easy	85.40 %	± 0.72 %	86.11 %	± 0.71 %
ARC-Challenge	62.54 %	± 1.41 %	66.21 %	± 1.38 %
HellaSwag	63.17 %	± 0.48 %	83.04 %	± 0.37 %
Winogrande	76.24 %	± 1.20 %	–	–

Readability (per age group)

• Flesch Reading Ease (German):

  • Grades 1–2: 77.99
  • Grades 3–4: 75.80
  • Grades 5–6: 76.84 → All levels exceed the 70-point threshold, ideal for CEFR A1–A2 comprehension.

• Wiener Sachtextformel:

  • Grades 1–2: 3.13
  • Grades 3–4: 3.42
  • Grades 5–6: 3.28 → Well below the threshold of 5, indicating age-appropriate readability.

Bias Analysis

• GenBit Score (target < 0.70–0.90):

  • Grades 1–2: 0.55
  • Grades 3–4: 0.54
  • Grades 5–6: 0.50 → Indicates low systemic gender bias.

• Female-to-Male Representation:

  • Grades 1–2: 59.26 % female
  • Grades 3–4: 52.98 % female
  • Grades 5–6: 65.82 % female → Slight female overrepresentation, contextually explainable, remains within acceptable bounds.

Summary

The LORA model meets or exceeds all Minimum Performance Requirements (MPR) across:

• Readability: Suitable for ages 6–12 based on Flesch and Wiener metrics.
• Fairness: Low gender bias with near-balanced representation.
• Robustness: Stable lexical and statistical behavior.

Evaluation results validate the model’s effectiveness as an educational language model for children, supporting safe and inclusive content generation.

Recommendations

• Content is optimized for German-language learning environments.
• Outputs should still be reviewed in production systems targeting children.
• Prompt structure should be followed for consistent output quality.

How to Get Started with the Model

from transformers import AutoModelForCausalLM, AutoTokenizer, pipeline

tokenizer = AutoTokenizer.from_pretrained("oscarstories/lorastral24b_0604")
model = AutoModelForCausalLM.from_pretrained("oscarstories/lorastral24b_0604")

generator = pipeline("text-generation", model=model, tokenizer=tokenizer)

prompt = "Schreibe eine Geschichte über das Mädchen namens Emma. Sie liebt Pferde. Erkläre den Begriff Demokratie. ############### Schulstufe 3-4"
output = generator(prompt, max_new_tokens=200, temperature=0.7, top_p=0.9, do_sample=True)

print(output[0]["generated_text"])

Training Details

Training Data

Our dataset is a carefully curated collection of high-quality, child-appropriate content designed specifically for developing language models for young audiences. This documentation outlines our data collection process, cleaning pipeline, and GDPR compliance measures.

Our corpus consists of two main public sources:

1. Klexikon

• Type: German-language children's encyclopedia
• Audience: Children aged 6–12
• License: Creative Commons Attribution-ShareAlike 4.0 (CC BY-SA 4.0)

Klexikon is a well-established German online encyclopedia created specifically for children. Unlike open-edit platforms, Klexikon enforces strict editorial guidelines and article reviews to ensure all content is suitable for young readers. Articles are written in simple, educationally aligned language and cover a wide range of school-relevant topics.

Klexikon is publicly recognized as a reliable educational resource, including endorsements by the German Federal Ministry of Family Affairs.

2. KiwiThek

• Type: Austrian children's encyclopedia and learning wiki
• Audience: Primary school learners and German language beginners
• License: Creative Commons Attribution-ShareAlike 4.0 (CC BY-SA 4.0), unless otherwise noted

KiwiThek is a project of the Wiener Bildungsserver (Vienna Education Network). It features articles in both standard German and “Deutsch einfach” — a simplified German variant tailored for children and language learners. This approach enhances accessibility and makes the content especially suitable for AI systems focused on inclusive learning.

Preprocessing

• Readability metrics vocabulary simplification
• Positive tone and emotional safety checks
• Balanced character/gender representation

Hyperparameters & Setup

• Precision: bf16 mixed
• Hardware: 2 × NVIDIA H100 (80 GB HBM3)
• VRAM used: 47.14 GB
• Throughput: ~7,100 tokens/sec
• Training duration: ~100 GPU hours

Environmental Impact

• Training used ~100 hours on 2 × H100 GPUs (16-bit precision)
• Throughput-optimized to reduce idle cycles per token

Authors and Contact

• Lead Authors: Dima Rubanov, Matthias Neumayer, Andreas Schaubmaier
• Organization: HeyQQ GmbH
• Contact: hello@oscarstories.com
• Project Website: https://oscarstories.com

---