Skip to content

[DMP 2026]: Developing a Cost-Efficient AI Model for Evaluating 21st Century Skills #2

@manua-glitch

Description

@manua-glitch

Ticket Contents

Description

TAP currently uses advanced proprietary AI models (e.g., Gemini) to evaluate student-submitted artifacts such as drawings, prototypes, and written responses against rubric-based frameworks measuring 21st-century skills like creativity, critical thinking, problem-solving, and agency. While effective, the cost per evaluation is high and becomes prohibitive at the scale of millions of government school students. This project aims to fine-tune an open-source Vision Language Model (e.g., LLaMA or similar) to replicate this evaluation capability at significantly lower cost, targeting below ₹0.10 per assessment.

Goals & Mid-Point Milestone

Goals

  • Clean, structure, and label TAP's existing dataset of student artifacts and rubric-based evaluations for model training
  • Research, select, and benchmark suitable open-source Vision Language Models for rubric-based evaluation tasks
  • Develop a fine-tuning pipeline optimized for rubric-based scoring accuracy
  • Benchmark the fine-tuned model against current Gemini-based evaluation and human evaluators
  • Achieve a cost per evaluation below ₹0.10 through efficient model architecture and deployment
  • [Goals Achieved By Mid-point Milestone]: Dataset preparation complete and initial fine-tuning pipeline running with preliminary benchmark results

Setup/Installation

No response

Expected Outcome

A fine-tuned open-source model capable of evaluating student artifacts
Cost-efficient inference pipeline
Evaluation benchmarks against existing systems
Documentation of training and deployment pipeline

Acceptance Criteria

No response

Implementation Details

  • Dataset: TAP's existing collection of student artifacts (images/videos) paired with rubric-based evaluations
  • Model: Open-source Vision Language Models such as LLaMA or similar architectures
  • Frameworks: PyTorch / TensorFlow for training and fine-tuning
  • Fine-tuning approach: Supervised fine-tuning optimized for structured rubric-based scoring
  • Cost optimization techniques: model quantization, efficient inference architecture
  • Benchmarking pipeline comparing model outputs with human evaluators and Gemini

Mockups/Wireframes

No response

Product Name

Open-Source AI Model for 21st Century Skills Assessment

Organisation Name

The Apprentice Project

Domain

⁠Education

Tech Skills Needed

Artificial Intelligence, Machine Learning, Computer Vision, Natural Language Processing, Python

Mentor(s)

TBD

Category

Machine Learning

Metadata

Metadata

Assignees

No one assigned

    Labels

    Type

    No type

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions