Inspiration

English essay writing improves through good feedback, but marking essays properly takes time. Teachers must judge whether a student answered the question, used the correct format, organised ideas clearly, and wrote accurate English.

SabiEssay was inspired by the need to make this feedback more accessible to Nigerian students and teachers, especially where internet access is unreliable or cloud AI tools are too expensive.

What it does

SabiEssay is an offline AI examiner for WAEC-style English compositions. Given the original question and a student’s essay, it returns:

  • Content: /10
  • Organisation: /10
  • Expression: /20
  • Mechanical Accuracy: /10
  • Total: /50
  • Short, actionable feedback

It runs locally on an ordinary 8 GB laptop without a discrete GPU or internet connection during inference.

How we built it

We studied 166 official composition questions from 34 WAEC examination sittings between 2008 and 2023, along with public Chief Examiner observations and detailed marking guides.

From this research, we created a controlled data-generation pipeline covering letters, articles, speeches, debates, reports, narratives and expository writing. Each synthetic essay was independently marked, checked against the rubric, and either accepted, flagged for review or rejected.

We fine-tuned Qwen3.5-2B to produce short, consistent examiner-style assessments. We then merged and quantized the model into a 1.27 GB Q4_K_M GGUF for CPU-only inference through llama.cpp.

On our laptop-like Ubuntu VM, the final model achieved:

  • 22.12 generated tokens per second
  • 1.95 GiB peak memory
  • Sperf: 100.00
  • Seff: 72.20

Challenges and lessons

The hardest challenge was teaching a small model that fluent writing is not automatically a good answer. An essay can sound impressive but still miss part of the question or use the wrong format.

We also learned that data quality mattered more than simply generating more examples. Weak labels caused the model to over-score poor essays, so we introduced independent marking, adjudication, validation rules and an accepted-only public dataset.

SabiEssay is not an official WAEC examiner and should not replace a teacher. It is designed as a private, offline first-pass assessment tool that helps teachers work faster and gives students clearer guidance.

Links

Built With

  • gemini-api
  • gguf
  • google-cloud-compute-engine
  • hugging-face-transformers
  • llama.cpp
  • lora
  • modal
  • peft
  • python
  • pytorch
  • qwen3.5-2b
  • unsloth
Share this project:

Updates

Submission history