How xrwr/kaggle-nvidia-nemotron-model-reasoning-challenge-1st-place works
This repository reproduces a competition solution: it derives synthetic reasoning problems and solver-generated traces from Kaggle training data, fine-tunes a Nemotron LoRA adapter, and evaluates or packages the result. The graph emphasizes that workflow and the separate local evaluation viewer. Most workflow links are documented in the README; source excerpts and the SOURCE INDEX support the solver, trace, training, and viewer relationships. Unsampled implementation details are represented only at their documented subsystem level.
Main entry points: sft.py SFT pipeline · __init__.py Solver dispatch · generate_reasoning_dataset.py Generate traces · create_dataset_split.py Create split
Components 21
Data preparation
- Create split
scripts/create_dataset_split.py - Dataset splitting
src/nvidia_nemotron_model_reasoning_challenge/dataset_split.py - Build priors
scripts/analysis/collect_problem_analysis_stats.py - Problem analysis
src/nvidia_nemotron_model_reasoning_challenge/problem_analysis - Equation prior
src/nvidia_nemotron_model_reasoning_challenge/problem_generation/problem_types/equation_transformation_prior.py
Problems and traces
- Generate problems
scripts/generate_problems.py - Synthetic generation
src/nvidia_nemotron_model_reasoning_challenge/problem_generation - Generate traces
scripts/generate_reasoning_dataset.py - Solver dispatch
src/nvidia_nemotron_model_reasoning_challenge/solver/__init__.py - Rule solvers
src/nvidia_nemotron_model_reasoning_challenge/solver - Trace generation
src/nvidia_nemotron_model_reasoning_challenge/generate/easy_reasoning.py - SFT trace sets
scripts/tools/merge_sft_jsonl_sources.py
Training and evaluation
- Train adapter
scripts/train_sft.py - SFT pipeline
src/nvidia_nemotron_model_reasoning_challenge/sft.py - Submission packaging
scripts/tools/package_kaggle_submission.py - Model evaluation
src/nvidia_nemotron_model_reasoning_challenge/eval/generic.py - Solver evaluation
scripts/evaluate_rule_solvers.py
Result inspection
- Evaluation viewer
app/evaluation_viewer.py - Viewer data access
src/nvidia_nemotron_model_reasoning_challenge/viewer_data.py
Other components
- Kaggle data
- Nemotron model
Connections 19
- Kaggle data → Create split: provides CSV
README.md - Create split → Dataset splitting: creates split
README.md - Kaggle data → Build priors: provides CSV
README.md - Build priors → Problem analysis: collects statistics
scripts/analysis/collect_problem_analysis_stats.py - Problem analysis → Equation prior: builds prior
README.md - Generate problems → Synthetic generation: generates problems
README.md - Equation prior → Synthetic generation: guides sampling
src/nvidia_nemotron_model_reasoning_challenge/problem_generation/problem_types/equation_transformation_generator.py - Generate traces → Trace generation: generates traces
README.md - Trace generation → Solver dispatch: requests solutions
src/nvidia_nemotron_model_reasoning_challenge/generate/easy_reasoning.py - Solver dispatch → Rule solvers: dispatches to
src/nvidia_nemotron_model_reasoning_challenge/solver/__init__.py - Generate traces → SFT trace sets: builds trace sets
README.md - SFT trace sets → Train adapter: supplies training data
README.md - Train adapter → SFT pipeline: runs training
scripts/train_sft.py - SFT pipeline → Nemotron model: fine-tunes
README.md - SFT pipeline → Submission packaging: produces checkpoint
README.md - Nemotron model → Model evaluation: generates answers
README.md - Solver evaluation → Solver dispatch: evaluates solvers
README.md - Model evaluation → Viewer data access: writes results
src/nvidia_nemotron_model_reasoning_challenge/eval/generic.py - Evaluation viewer → Viewer data access: loads results
app/evaluation_viewer.py