QRPO Paper: Llama SFT LeetCode Sandbox Reference Completions and Rewards

A dataset from the huggingface platform, created by author skandermoalla and last updated on December 8, 2025. It contains reference completions and rewards for a specific model and reward model, intended for training with the QRPO reference codebase. This collection supports the paper 'Quantile Reward Policy Optimization: Alignment with Pointwise Regression and Exact Partition Functions'.

Use Cases

• Training reinforcement learning agents for code generation based on reference completions and reward signals.
• Benchmarking off-policy optimization algorithms based on the provided reward model outputs.
• Reproducing experiments from the QRPO research paper based on the linked dataset collection.
• Fine-tuning language models on programming tasks using structured reward feedback.

Strengths

• Dataset is explicitly linked to a published research paper (arXiv:2507.08068).
• Includes reference completions and rewards, which are key components for reinforcement learning from human feedback (RLHF) workflows.
• Last update timestamp (2025-12-08 13:27:14) is provided.

Limitations

• Column names, data types, and sample rows are unknown, requiring inspection after download.
• The total number of rows, file formats, and license information are not specified.
• Data may reflect bias inherent to the specific model, reward model, and sandbox environment used for generation.

Provenance

Source

huggingface

Collection Method

Likely generated from a specific model and reward model within a LeetCode sandbox environment, as part of research for the QRPO paper.

Time Range

null

Freshness

Last updated 2025-12-08 13:27:14.

Geography

null