How manavsharma2707/trace-and-ace-comp works
Trace the Ace is a DrivenData competition solution for estimating a student’s probability of answering the next assessment question correctly. Its documented workflow develops transcript features and cross-validation folds, trains feature-based and LLM models, then combines their predictions in a packaged inference submission. The README also documents leakage-safe objective encoding and an embedding-kNN fallback. Source files were not sampled, so links represent documented relationships only; implementation details inside submission.zip are not independently verified.
Main entry points: submission.zip Packaged inference · 01-eda-features.ipynb Feature engineering · 03-llm-finetune.ipynb Qwen classifier · assets Model assets
Components 16
Data preparation
- Feature engineering
notebooks/01-eda-features.ipynb - Transcript features
notebooks/01-eda-features.ipynb - Grouped CV folds
notebooks/01-eda-features.ipynb
Model development
- Logistic baseline
notebooks/02-baseline-model.ipynb - LLM fine-tuning
notebooks/03-llm-finetune.ipynb - LightGBM model
submission.zip - Objective encoding
submission.zip - Embedding-kNN fallback
submission.zip - Qwen classifier
notebooks/03-llm-finetune.ipynb
Inference and blending
- Packaged inference
submission.zip - Weighted blend
submission.zip - Probability prediction
submission.zip
Packaged artifacts
- Model assets
assets - Submission package
submission.zip
Other components
- Competition data
- Competition runner
Connections 16
- Competition data → Feature engineering: provides data
- Feature engineering → Transcript features: creates features
- Feature engineering → Grouped CV folds: creates folds
- Transcript features → Logistic baseline: trains baseline
- Transcript features → LightGBM model: trains model
- Grouped CV folds → LightGBM model: guides validation
- Objective encoding → LightGBM model: encodes objectives
- Embedding-kNN fallback → Objective encoding: handles unseen
- LLM fine-tuning → Qwen classifier: trains classifier
- Competition runner → Packaged inference: invokes
- Submission package → Packaged inference: packages
- Model assets → Packaged inference: supplies weights
- Packaged inference → LightGBM model: predicts features
- Packaged inference → Qwen classifier: predicts transcript
- Packaged inference → Weighted blend: combines predictions
- Weighted blend → Probability prediction: returns probability