Deploy quantized adapted small models offline with a teacher in the loop rather than as autonomous graders¶

Strategy · All strategies
Evidence · no claims cited

Description¶

The article recommends deploying the adapted, quantized 4B model as an offline binary for local inference in resource-constrained environments, but explicitly positions it as an assistive tool: PhysicsMate and derived models "should be placed as a tool to assist teaching, reviewed by a teacher rather than deployed as an autonomous grader". Because the best offline model reaches only 50.9% closed-book accuracy, direct unmoderated deployment to students risks reinforcing misconceptions, so a teacher-in-the-loop framework is advised for classrooms with limited connectivity.

Design Implications¶

Context¶

Requirements¶

• At least one device capable of handling the content must be present in the classroom, and public release of derived materials must include a provenance statement and adhere to NCTB licensing terms.

Constraints¶

• The best offline model still has substantial room for error at 50.9% closed-book accuracy; offline deployment does not eliminate the resource gap.

Target Learners¶

• Secondary students in resource-constrained, limited-connectivity classrooms in South Asia

Target Learning Goals¶

• Curriculum-aligned physics support without reliable internet connectivity

Related Strategies¶

• Evaluate quantisation choices empirically on the target task and hardware rather than inheriting them from other studies
• Use pedagogical knowledge benchmarks to guide model selection and fine-tuning for educational AI, with ethical guardrails and human-in-the-loop deployment
• Deploy fine-tuned compact local 8B LLMs as free classroom and home story generators with adult human oversight and safety filtering
• Disclose quantization trade-offs, audit bias, and establish data governance before live deployment of quantized educational AI
• Integrate AI-formative tools state-wide and deploy offline-capable Edge AI servers

Examples¶

-

Key Sources¶

• Rashid Azraf Jahin, Saadman Sajid, Khan Raiyan Ibne Reza, Sumaiya Tabassum Nimi. (2026). PhysicsMate: A Curriculum-Grounded Bengali Benchmark for Secondary Physics QA with Small-Model Adaptation. IEEE CSDE 2026. https://arxiv.org/abs/2610.00664