Flawed Positive Benchmarks for Reasoning Model Training

FAPO Critic contains constructed benchmark data for training a generative reward model. The dataset was created by author dyyyyyyyy for the FAPO research project and was last updated on the platform in October 2025. It is sourced from ProcessBench and forms the FlawedPositiveBench used to train the FAPO-GenRM-4B model.

Use Cases

• Training a generative reward model (FAPO-GenRM-4B) to score reasoning steps.
• Evaluating policy optimization algorithms for reliability using constructed benchmark queries and responses.
• Analyzing model performance on flawed-positive examples to improve reasoning safety.

Strengths

• Data is specifically constructed for a published research method (FAPO).
• Dataset supports training of a named 4-billion-parameter model (FAPO-GenRM-4B).

Limitations

• Specific row count, column names, and data size are unknown.
• The dataset's scope is narrowly focused on a single research methodology, limiting generalizability.

Provenance

Source

ProcessBench, as part of the FAPO research project.

Collection Method

Constructed benchmark data (FlawedPositiveBench).

Time Range

null

Freshness

Last updated on the platform in October 2025.

Geography

null