Skip to content

v1.5.0: Any Post-Training

Choose a tag to compare

@warlockee warlockee released this 27 Feb 01:13
· 53 commits to main since this release

17 Algorithms — Any Post-Training

oxRL now ships 17 algorithms covering the full post-training landscape.

New in v1.5.0

Supervised Learning (12 algorithms):

Algorithm Description
SFT Supervised Fine-Tuning
DPO Direct Preference Optimization
ORPO Odds Ratio Preference Optimization
KTO Kahneman-Tversky Optimization
CPT Continued Pre-Training (all-token loss)
KD Knowledge Distillation (CE + KL)
RM Reward Model Training (Bradley-Terry)
OnlineDPO DPO with online rejection sampling
RFT Rejection Sampling Fine-Tuning
SPIN Self-Play Improvement
IPO Identity Preference Optimization (squared loss)
SimPO Simple Preference Optimization (reference-free)

Reinforcement Learning (5 algorithms):

Algorithm Description
SGRPO Stable GRPO with clipped surrogate loss
CISPO Clipped importance-sampling policy optimization
PPO Proximal Policy Optimization with GAE + value clipping
RLHF GRPO with trained reward model
RLAIF GRPO with AI-generated rewards

Other Changes

  • oxrl/algs/__init__.py now exports all 14 algorithm classes
  • 9 new example configs in registry/examples/
  • README updated: honest VLM support (VLA claim removed), complete algorithm tables
  • 64 tests passing (was 33)
  • New config fields: simpo_gamma, ppo_vf_clip, ppo_tau, ppo_gamma, kd_alpha, kd_temperature, reward_threshold, reward_model_path

Breaking Changes

  • oxrl/algs/simpo.py rewritten from Ray actor to SL-compatible DeepSpeed algorithm
  • oxrl/algs/ppo.py refactored to self-initialize (like GRPO)