Downloads · 30 days
0
mathlingua-spec
cosmicmicra/mathlingua-spec
mathlingua-spec is a machine learning model from cosmicmicra. Use it for the machine learning task on the model card, and read the license before you ship it in a product.
An adaptive tutoring system for Spanish-speaking students (grades 6–8) transitioning to English-medium mathematics education. Uses progressive bilingual scaffolding and a hybrid adaptive algorithm to personalize diffi…
Access
Public
Updated May 3, 2026
Repo size
—
Likes
0
Public
Repo makeup
Click a slice to open those files.
.md106 KB · 69%
At a glance
- Access
- Public
- Created
- Apr 27, 2026
- Updated
- May 3, 2026
- SHA
- 5dcf37af
Model card
From the Hugging Face model README
MathLingua — Adaptive Bilingual Math Tutoring System
An adaptive tutoring system for Spanish-speaking students (grades 6–8) transitioning to English-medium mathematics education. Uses progressive bilingual scaffolding and a hybrid adaptive algorithm to personalize difficulty.
📄 Documents
| File | Description |
|---|---|
| technical_specification.md | Full paper-ready technical specification (~60KB, 1200+ lines) |
| system_architecture.md | Detailed system architecture with component diagrams, data flows, API contracts |
🏗️ Key Components
Four-Level Scaffolding
- L1: Simplified English (shorter sentences, simpler vocabulary)
- L2: Bilingual keywords inline (English with Spanish annotations)
- L3: Full Spanish translation
- L4: Step-by-step solution reveal
Hybrid Adaptive Algorithm
- Elo Rating: Overall student ability tracking with hint-weighted outcomes (1.0/0.75/0.50/0.25/0.0)
- Bayesian Knowledge Tracing (BKT): Per-topic mastery estimation with slip adjustment for scaffold usage
- Thompson Sampling: ZPD-constrained question selection with exploration/exploitation balance
Engineered Features
- Language Dependency Score (LDS): Quantifies reliance on linguistic scaffolding (0 = English-independent, 1 = Spanish-dependent)
- Math Confidence Score (MCS): Quantifies mathematical ability independent of language (0 = struggling, 1 = confident)
Technology Stack
- Frontend: Next.js 14+ / TypeScript / Tailwind CSS
- Backend: Firebase (Auth, Firestore, Cloud Functions)
- LLM V1: Google Gemini 2.0 Flash
- SLM V2: Qwen2.5-3B-Instruct (QLoRA fine-tuned)
- Adaptive Engine: Client-side TypeScript (zero-latency decisions)
📊 Difficulty Taxonomy
15 sub-levels across 3 tiers, ordered by linguistic complexity (not mathematical difficulty):
- Level 1 (Easy): 1.1–1.5, FK grade 1.0–6.0
- Level 2 (Medium): 2.1–2.5, FK grade 5.5–10.0
- Level 3 (Hard): 3.1–3.5, FK grade 9.5–14.0
Validated with monotonically increasing Flesch-Kincaid readability scores across all 15 levels.
📝 License
This project documentation is provided for educational and research purposes.
- Created
- Apr 27, 2026
- Updated
- May 3, 2026