---
pretty_name: SelectiveRM benchmark embeddings
tags:
- reward-modeling
- optimal-transport
- helpsteer
- ultrafeedback
- safe-rlhf
---

# SelectiveRM benchmark embeddings

Cached frozen embeddings produced from HelpSteer, UltraFeedback, and
PKU-SafeRLHF with `sfairXC/FsfairX-LLaMA3-RM-v0.1`. HelpSteer's `test`
artifact maps to its official `validation` split. Each `binary_noisy` artifact
contains the exact 20% symmetric-noise data used for the symmetric versus
prediction-preserving partial-OT comparisons. Training uses seed 42; the
repository's noise simulator uses its separately fixed seed 0.

See `manifest.json` for file sizes, SHA-256 checksums, source identifiers, and
noise configuration. Safetensors should only be loaded from a trusted source
and must be validated before training.