MyVillage Activity Type Classifier
Classifies an activity into one of REFLECTION, RESEARCH, COLLABORATE, CREATE, PRACTICE, EXPERIENCE, or TEACH. This is single-label classification with softmax.
Source code, training instructions, and API: https://github.com/Jvmes3/Activity-Type-Classification-Model
Input format
Use the same formatting as training: normalize whitespace within each field and join these three lines, retaining empty fields when omitted:
title: Learning journal
description: Review your learning.
instructions: Explain what changed in your understanding.
Load this repository with AutoTokenizer.from_pretrained and
AutoModelForSequenceClassification.from_pretrained, tokenize with truncation and
max_length=512, and apply softmax to the logits. Map the highest-scoring index
through model.config.id2label. The GitHub project's ActivityClassifier does this.
Evaluation and limitations
See evaluation.json, when present, for metrics, class counts, the source-data hash,
and base model. Metrics depend on the uploaded checkpoint and dataset; the repository
name alone does not establish training quality. Scores from the included synthetic
fixture are pipeline demonstrations, not evidence of real-world accuracy.
Confidence scores are uncalibrated. Review mixed-objective activities and evaluate on independent, approved MyVillage data before using predictions for routing. Inputs longer than 512 tokens are truncated. Confirm training-data provenance and usage rights before distributing a checkpoint.
- Downloads last month
- 33