Baileyy/stepwise-ielts-modernbert-moe

Stepwise IELTS ModernBERT MoE

This is the frozen fine-tuned encoder and custom mixture-of-experts regression head used in the matched experiment and deployed by Stepwise. All model tensors and the tokenizer are included. It takes an IELTS Writing Task 2 question and essay and estimates four half-band scores: Task Response (TR), Coherence and Cohesion (CC), Lexical Resource (LR), and Grammatical Range and Accuracy (GRA). No requested target band is supplied to the scorer.

Evaluate and reproduce

We trained a new four-expert head for 12 epochs with the encoder frozen, then trained the encoder and head jointly for three epochs on the same 4,096 essays. The shared 256-essay development split selected joint epoch 2. The test registered 72 essays; every model scored 70 in common. Results below are mean absolute error (IELTS band points, lower is better) against Writing9 site ratings.

Overall	TR	CC	LR	GRA
1.079	1.086	1.100	1.000	1.129

Full benchmark, including Gemini with and without RAG · Companion model.

Load and score

git clone https://huggingface.co/Baileyy/stepwise-ielts-modernbert-moe
cd stepwise-ielts-modernbert-moe
python -m pip install -r requirements.txt
python score.py example.json

Install Git LFS before cloning, or use hf download Baileyy/stepwise-ielts-modernbert-moe --local-dir stepwise-ielts-modernbert-moe. To score a different essay, supply a JSON file with string fields prompt and essay. The bundled example is a Stepwise demonstration, not a training or evaluation record. This is a numeric scorer with a custom head; the base-model text-generation or fill-mask pipeline does not run it.

Limits and provenance

Practice and research estimates only. Writing9 site scores were not independently verified by IELTS examiners, and performance on the matched sample does not establish examiner accuracy. The data source's listed license is unknown; this repository does not redistribute raw essays or their ratings.

Base: answerdotai/ModernBERT-base, Apache 2.0. Stepwise trained the scoring components described above. See checkpoint.json for hashes and aggregate results.