retrain 0.2.0

RLVR training framework for LLMs

retrain

RLVR (Reinforcement Learning with Verifiable Rewards) training framework for LLMs. Train reasoning models on MATH with composable advantage functions and adaptive scheduling.

Features

• Composable advantages -- GRPO, MaxRL, GTPO entropy weighting, HICRA planning amplification, SEPA entropy pooling
• Pluggable inference -- PyTorch, MAX, vLLM, SGLang, or any OpenAI-compatible server
• Pluggable rewards -- string match, symbolic math, LLM judge, or bring your own
• Back pressure -- USL+Roofline adaptive batch sizing
• Campaign orchestrator -- sweep all 5 conditions across seeds with wandb logging
• Checkpoint resume -- save and restore full trainer state across preemptions

Install

pip install -e .

Quick start

# 1. Drop a config
cp retrain.toml my_run.toml

# 2. Train
retrain my_run.toml

# 3. Override from CLI
retrain my_run.toml --seed 42 --wandb-project my-project

Configuration

All configuration lives in a TOML file. See retrain.toml for the default config, or run retrain help for the full reference.

Documentation

Full documentation: retrain.readthedocs.io

• Getting Started
• Configuration Reference
• Advantage Functions
• SEPA Scheduling
• Reward Functions
• Inference Engines