OneRec 2502.18965 Reproduction Scaffold (PyTorch)

This repository now contains a minimal, runnable PyTorch scaffold for reproducing OneRec (arXiv:2502.18965) in a disciplined, iteration-friendly way.

What is included

• Structured paper-to-code plan in docs/spec_2502.md
• Minimal sequential recommendation training loop (BPR objective)
• Config-driven entrypoint
• Core ranking metrics: Recall@K and NDCG@K
• Synthetic data path so the pipeline can run end-to-end immediately

Note: this is a starting point focused on engineering reproducibility. Exact paper-level alignment still requires dataset/protocol implementation from the paper.

Quick start

python train.py --config configs/onerec_2502_minimal.yaml

Project structure

• configs/ - experiment configs
• docs/ - structured reproduction specification
• docs/results/ - experiment result reports
• onerec/ - package modules
  • data/ - dataset and dataloader helpers
  • models/ - model definitions
  • training/ - trainer logic
  • metrics/ - ranking metrics

Next steps

1. Implement exact dataset split protocol from 2502.18965.
2. Replace synthetic dataset path with real benchmark pipeline.
3. Add paper-specific model blocks and losses as ablations.
4. Run multi-seed experiments and compare against reported metrics.

Results

• ML-1M (full, all-item eval, MPS): 2026-02-18-ml1m

Dashboard

• Start the dashboard server:
  • . .venv/bin/activate
  • python dashboard.py
• Open http://localhost:8765/ to view loss/time/Recall@K/NDCG@K

Dataset download links

If you want to manually download paper-related recommendation datasets first, see:

• docs/dataset_links.md