TEKS Classifier

Complete
High-school CS project
Python · DistilBERT · Hugging Face Transformers +7
2024–2025

A lesson-plan review pipeline that fetches Canvas assignments, predicts Texas US History standards with a fine-tuned DistilBERT model, and writes ranked suggestions to Google Sheets. The original automation ran nightly on GitHub Actions from February to October 2025.

Python
DistilBERT
Hugging Face Transformers
AutoTrain
PyTorch
BeautifulSoup
Canvas REST API
Google Sheets API
GitHub Actions
pytest

GitHub Model on Hugging Face

Overview

I built this for a teacher who needed to connect lesson plans to Texas's US History standards. It reads the plans already in Canvas and puts three ranked suggestions in a shared Google Sheet for review, rather than asking the teacher to move the lessons into another tool.

The work covers model selection and fine-tuning, lesson-plan parsing, Canvas and Sheets integration, nightly automation, retry-safe state tracking, and evaluation on real lessons. The suggestions are a first pass for a teacher, not an automatic standards-alignment decision.

What I built

Nightly sync: the Google Sheet records which assignments have been processed, while new Canvas lessons are parsed, classified, and appended in a batch.

Model selection and training: after a larger Llama fine-tuning attempt ran out of GPU memory, I switched to CPU-capable DistilBERT. I fine-tuned and published the AutoTrain checkpoint, and added a local Transformers Trainer setup for the current dataset.
Lesson parsing: built a heading-aware BeautifulSoup parser for inconsistent Canvas HTML. It keeps objectives, questions, and practice activities, while excluding logistics and teacher-cited standards. Tests distinguish headings from inline bold text and standard codes.
Inference: return the top three labels with softmax model probabilities, using the 128-token input limit from training, so the teacher has a ranked shortlist to review.
API integration and automation: fetch Canvas assignments and append new predictions to Sheets with a service account. The current design uses the sheet's assignment IDs as its only progress state and compares IDs as strings, fixing the integer/string mismatch that caused duplicate uploads in the original JSON-file approach.
Operational tooling: added a dry-run CLI, environment-based configuration, credential exclusions, and fast parser/pipeline tests in GitHub Actions that do not need to download the model.

Evaluation on real lessons

Teachers often cite standards in the lesson itself. I built an extractor for those citations and used them as reference labels for evaluation, while withholding the citation section from the model's input. That let me measure performance on real lessons without creating a new hand-labeled dataset.

Evaluation separates instructional text from the teacher's cited standards, then compares predictions against those held-out citations.

Training-style validation: the AutoTrain checkpoint recorded 53.7% top-1 accuracy on held-out paraphrases of standards.
Real lesson plans: on 26 lessons evaluated against teacher-cited standards, top-1 was 23% and top-3 was 27%. The corresponding random baselines were 2% and 6%. This is a small evaluation set, and its scores are separate from the training-style result.
Coverage limit: the current training CSV contains 537 examples across 107 standards. The published checkpoint predates the last 13 standards and cannot predict them; the local training script supports the full current dataset.

What I learned

The main limitation is the gap between short training paraphrases and real lesson plans. Filtering the input helps, but does not replace training on actual lessons. The evaluation pipeline already extracts teacher-cited examples for a future training set. A multi-label model is another next step, since one lesson can cover several standards.

The deployment work mattered as much as the model: choosing something small enough for a free CPU runner, keeping credentials out of the repository, and making a nightly job safe to rerun turned a classifier experiment into a usable review workflow.

Overview

What I built

Nightly sync: the Google Sheet records which assignments have been processed, while new Canvas lessons are parsed, classified, and appended in a batch.

Model selection and training: after a larger Llama fine-tuning attempt ran out of GPU memory, I switched to CPU-capable DistilBERT. I fine-tuned and published the AutoTrain checkpoint, and added a local Transformers Trainer setup for the current dataset.

Lesson parsing: built a heading-aware BeautifulSoup parser for inconsistent Canvas HTML. It keeps objectives, questions, and practice activities, while excluding logistics and teacher-cited standards. Tests distinguish headings from inline bold text and standard codes.

Inference: return the top three labels with softmax model probabilities, using the 128-token input limit from training, so the teacher has a ranked shortlist to review.

API integration and automation: fetch Canvas assignments and append new predictions to Sheets with a service account. The current design uses the sheet's assignment IDs as its only progress state and compares IDs as strings, fixing the integer/string mismatch that caused duplicate uploads in the original JSON-file approach.

Operational tooling: added a dry-run CLI, environment-based configuration, credential exclusions, and fast parser/pipeline tests in GitHub Actions that do not need to download the model.

Evaluation on real lessons

Evaluation separates instructional text from the teacher's cited standards, then compares predictions against those held-out citations.

Training-style validation: the AutoTrain checkpoint recorded 53.7% top-1 accuracy on held-out paraphrases of standards.

Real lesson plans: on 26 lessons evaluated against teacher-cited standards, top-1 was 23% and top-3 was 27%. The corresponding random baselines were 2% and 6%. This is a small evaluation set, and its scores are separate from the training-style result.

Coverage limit: the current training CSV contains 537 examples across 107 standards. The published checkpoint predates the last 13 standards and cannot predict them; the local training script supports the full current dataset.

What I learned

TEKS Classifier

Overview

What I built

Evaluation on real lessons

What I learned

Projects

TEKS Classifier

Overview

What I built

Evaluation on real lessons

What I learned