Research Models & Publications

APAS AI Lab develops lightweight, curriculum-aware AI models for Indian education and creativity.

EduQwen-1.8B: Curriculum-Aligned Tutor

Overview: A lightweight language model fine-tuned on NCERT, NEET, and JEE content to generate accurate, step-by-step educational explanations.

Base Model: Qwen-1.8B-Chat (TheBloke / Qwen GGUF)

Intended Use:

  • Explain physics derivations and chemistry mechanisms
  • Generate NEET-aligned practice questions
  • Provide doubt-solving in Hinglish

Training Data (Planned): NCERT textbooks (Class 11–12), Irodov problem sets, verified solutions from APAS Learners.

Evaluation: Accuracy on 100 NEET 2025 problems, pedagogical clarity (faculty-rated).

Status: Fine-tuning in progress on RunPod (A10 GPU).

Pruna-Indian-Images: Culturally-Grounded Vision

Overview: A fine-tuned SDXL model that generates images reflecting Indian festivals, classrooms, and cultural contexts.

Base Model: Stable Diffusion XL 1.0 (stabilityai)

Intended Use:

  • Diwali/Eid/Holi promotional visuals
  • Classroom and student diversity scenes
  • Educational diagram generation

Training Data (Planned): Public-domain Indian festival imagery, NCERT illustrations, synthetic coaching visuals.

Evaluation: Cultural authenticity (human eval), prompt adherence (CLIP score).

Status: Dreambooth fine-tuning planned.

AnimateDiff-Indian: Short Cultural Animations

Overview: AnimateDiff-Lightning tuned for 3–5 second animations of Indian motifs (diyas, rangoli, books).

Base Model: AnimateDiff-Lightning (guoyww)

Intended Use:

  • Social media promo clips for coaching centers
  • Course intro animations
  • Festival greeting videos

Training Data (Planned): Short clips of festive elements, animated NCERT diagrams.

Evaluation: Motion smoothness (FVD), cultural relevance.

Status: Inference testing on RunPod.

HinglishQwen-1.8B: Indian Language Reasoning

Overview: Qwen-1.8B fine-tuned for Hinglish with subject-specific vocabulary (e.g., “acceleration”, “molarity”).

Base Model: Qwen-1.8B (TheBloke)

Intended Use:

  • Answer student doubts in Hinglish
  • Generate bilingual concept explanations
  • Create regional-language quizzes

Training Data (Planned): Hinglish coaching notes, YouTube educator transcripts, NCERT summaries.

Evaluation: Hinglish fluency (BLEU, human eval), subject term accuracy.

Status: Dataset curation in progress.

MathQwen-1.8B: Mathematical Reasoning

Overview: Qwen-1.8B enhanced with MathCoder techniques for step-by-step math problem solving.

Base Model: Qwen-1.8B + MathCoder prompt tuning

Intended Use:

  • Solve algebra, calculus, and word problems
  • Generate similar practice questions
  • Verify student answers

Training Data (Planned): RD Sharma, RS Aggarwal, JEE Main math questions (2010–2025).

Evaluation: Accuracy on JEE math section, step completeness.

Status: Prompt engineering phase.

All models are developed by APAS AI Lab — a research division of Apas Learning Private Limited (Startup India Recognized, Est. 2019).

Explore our products: apashub.com