paulokewunmi/SabiEssay-GGUF

SabiEssay GGUF

SabiEssay is a compact, offline assistant fine-tuned to provide first-pass assessment of WAEC-style English compositions. Given the original composition question and one complete student essay, it returns:

• Content /10;
• Organisation /10;
• Expression /20;
• Mechanical Accuracy /10;
• Total /50;
• concise feedback and improvement priorities.

It is designed for CPU inference on ordinary consumer laptops using llama.cpp. Internet access is not required after downloading the model.

Model details

Property	Value
Base model	Qwen/Qwen3.5-2B
Base revision	15852e8c16360a2fea060d615a32b45270f8a8fc
Adaptation	LoRA rank 16, alpha 16, all linear layers
Quantization	GGUF Q4_K_M
Effective parameters	1,881,825,088
GGUF size	1,274,396,032 bytes
GGUF SHA-256	f3a0f36e981b1660ed8147082cff60a6f218b2f9fa45ac6dfdbd74fc00504d83
Conversion runtime	llama.cpp commit b15ca938a
Language	English

The final artifact passed a Linux CPU-only llama-bench smoke test with four threads and -ngl 0.

Recommended prompt

Act as a WAEC English composition examiner. Assess the student composition using Content /10, Organisation /10, Expression /20, and Mechanical Accuracy /10. The Total /50 must equal the four component scores. Give concise evidence-based feedback.

QUESTION:
<original composition question>

STUDENT ESSAY:
<complete essay>

Use deterministic decoding, disable thinking, and keep responses concise. The training context was 2,048 tokens.

Run with llama.cpp

llama-cli \
  --model SabiEssay-Q4_K_M.gguf \
  --conversation \
  --reasoning off \
  --temperature 0 \
  --predict 128

Evaluation

On a 27-item synthetic held-out grading set, the full-precision adapter produced:

• 100% parseable component-score responses;
• total-score MAE of 2.59/50;
• median absolute total error of 1.5 marks;
• 15/27 totals within two marks of the synthetic adjudicated reference.

These are synthetic-reference results, not agreement measurements against official WAEC examiners.

Intended use

• first-pass feedback for teachers and learners;
• offline educational experimentation;
• research on compact rubric-aligned language models.

Limitations

SabiEssay is not a certified WAEC examiner and must not be represented as producing official results. It can over-score polished but irrelevant writing, invent errors in clean prose, repeat fields, or mishandle adversarial and instruction-injected essays. It is not reliable for exhaustive proofreading, comprehension, summary, literature, objective questions, multilingual assessment, or consequential decisions without human review.

Provide one conventional English composition at a time and retain a qualified teacher as the final decision-maker.

Training data

The public dataset linked above contains the 194 accepted synthetic records. The training run also used 92 additional synthetic records flagged for human review; those noisier records are deliberately not included in the public dataset preview.

Historical WAEC materials were used privately to study task structure and marking principles. Historical questions, marking guides, scans, and real student scripts are not distributed in either repository.

Attribution and disclaimer

SabiEssay was developed by Paul Okewunmi. It is an independent project and is not affiliated with, approved by, or endorsed by the West African Examinations Council.

The base model is Qwen3.5-2B, distributed under Apache-2.0. This repository distributes the fine-tuned GGUF under the same license. See LICENSE and NOTICE.