How aaronmueller/cs599-fall26-hw2 works
This repository is a post-training homework package, not a complete training application: the README describes an assignment and directs students to implement missing functions, while the sampled runtime code provides model loading, math-answer grading, and vLLM helpers. A typical intended workflow is to use prompt/data resources with a model and evaluate generated answers, but the shown files do not establish a training orchestrator or connect those resources to the helpers. The graph therefore shows only supported runtime relationships; dataset groupings reflect the tree, not inspected dataset contents.
Main entry points: vllm_utils.py vLLM helpers · checkpoint.py Model loader · drgrpo_grader.py Math grader
Components 11
Model and inference
- Model loader
student/checkpoint.py - Model checkpoint
- vLLM helpers
student/vllm_utils.py - vLLM server
Answer evaluation
- Math grader
student/drgrpo_grader.py
Homework datasets
- Math reasoning
data/gsm8k - Human preferences
data/hh - MMLU questions
data/mmlu - Safety examples
data/simple_safety_tests - Alpaca evaluations
data/alpaca_eval
Other components
- Assignment user
Connections 7
- Assignment user → Model loader: loads model
student/checkpoint.py - Model loader → Model checkpoint: loads weights
student/checkpoint.py - Assignment user → vLLM helpers: requests inference
- vLLM helpers → vLLM server: controls server
student/vllm_utils.py - vLLM helpers → vLLM server: requests completions
student/vllm_utils.py - vLLM helpers → vLLM server: syncs weights
student/vllm_utils.py - Assignment user → Math grader: grades answers