Essay and IELTS writing coach (Labs build #6, started 2026-10-26)

Paste an essay, get back: six skill scores, the argument parts tagged (lead, position, claims, evidence, counterclaim, rebuttal, conclusion), weak parts flagged, and a band. Build #5 on the agreed list; picked up after outfit swap was paused.

What the page does

You paste your essay (and optionally the question it answers).
Six scores from 1 to 5: cohesion, syntax, vocabulary, phraseology, grammar, conventions. Shown as a radar, with the weakest two called out in plain words.
The essay comes back with its argument parts coloured, and each part marked effective / adequate / ineffective. That is where the "weak sentences flagged" feedback comes from: an ineffective claim with no evidence after it, a missing position, no counterclaim, a lead that is one sentence long.
A band: an IELTS-style overall band for Task 2 essays. Honest about its limits (see data).

Data (all public, all human-scored except where said)

ELLIPSE corpus (Crossley et al.): ~6.5k essays by English language learners, grades 8-12, overall + the six analytic scores, CC BY-NC-SA 4.0. The Kaggle "Feedback Prize - English Language Learning" train set (3.9k) is a subset.
PERSUADE 2.0 corpus: 25k argumentative essays, holistic score 1-6, every argument part annotated with its type and its effectiveness. CC BY-NC-SA 4.0. The Kaggle Feedback Prize 2021 / Effectiveness / AES 2.0 data come from it.
IELTS: "IELTS Writing Scored Essays Dataset" (Kaggle, mazlumi), ~1,200 essays with an overall band and an examiner comment, licence "other", origin not stated. A second set (xntrng15) has criterion scores produced by GPT-5 Mini, so it is not ground truth; at most a sanity check. The band head is therefore small-data and will be labelled as an estimate on the page.
Licence consequence: CC BY-NC-SA means the model weights are shared under the same licence and the page is non-commercial. That is fine for Labs.

Recipe (from the Kaggle winners, cut down to one small GPU)

One backbone: deberta-v3-base (large if the budget allows), max length 1024, newline kept as a special token (every winner found DeBERTa's tokenizer drops "\n"; +0.01).
Head A, six scores: regression, MCRMSE, dropout off, layer-wise lr decay, AWP optional. Winners' single deberta-v3-large: ~0.45 MCRMSE; base ~0.46-0.47. Target: ~0.47 on a held-out ELLIPSE fold.
Head B, argument parts: token classification over the essay (BIO over 7 types), decode with "repair broken spans, one lead/position/conclusion per essay". Winners' single model ~0.70 F1 (span overlap); target ~0.65.
Head C, effectiveness: whole essay in once, special start/end token per part, one vector per part -> 3 classes (the "essay group model" every top team converged on). Target log loss ~0.60 single model.
Band: a small model on top of the six scores + part statistics + length, fit on the 1.2k IELTS essays with cross-validation; report the within-half-band accuracy honestly.
Budget: each head is one or two hours on an A10G at ~$1.10/h. Whole build well under $10 of the $30 on labs3.

Where things are

Code: this folder (coach/ model + data, scripts/ Modal jobs). Data: Modal volume essay-data (labs3).
Runs and weights: HF nonsodev/essay-coach (CC BY-NC-SA, public is fine).

Results

scores-v1 (2026-10-27): deberta-v3-base, 4 epochs, 24 min on an A10G. MCRMSE 0.446 on the 2,571 ELLIPSE test essays (per trait 0.41-0.48). The Kaggle ELL winners' single deberta-v3-large was ~0.45, so this is at that level.
Serving: scripts/serve.py is a CPU web endpoint on Modal (scales to zero); web/index.html takes ?api=.
parts-v1 (2026-10-27): deberta-v3-base token model, 3 epochs on 24k PERSUADE 2.0 essays, 103 min on an A10G. On 2,000 held-out essays: argument-part F1 0.656 (Feedback Prize 2021 metric; winners' single models ~0.70), effectiveness log loss 0.478 on gold parts (not the competition's test set, so only roughly comparable to their 0.57-0.60).
band-v1: ridge on the six raw scores + part counts + length, 5-fold CV on 1,435 IELTS essays: RMSE 0.79 bands, within half a band 67%, within one band 89%, correlation 0.66 (predicting the mean would be RMSE 1.06). Shown on the page as an estimate with that error rate.
Endpoint: https://sylvayoutube--essay-coach-server-read.us-east.modal.run (CPU, ~0.8 s per essay).