v0.4.4

ART 0.4.4 Release Notes

New Features

• SkyPilot Integration Enhancement: Added SkyPilot extras support for improved cloud deployment
  capabilities (#255)
• Reward System Improvements: Added experimental support to not scale rewards, providing more
  flexibility in reward configuration (#fd2a118)

Documentation & Examples

• New Tutorial: Added temporal-clue-7b.ipynb notebook demonstrating temporal reasoning capabilities
  (#a2802b3)
• Enhanced Documentation: Updated RULER documentation with comprehensive guidance on combining
  rewards (#250)
• ART•E Integration: Added ART•E notebook examples to documentation (#242, #240)

Bug Fixes & Improvements

• Dependency Management: Reverted to previous version of gql to resolve compatibility issues (#249)
• Unsloth Integration: Added experimental logprob pre-calculation support for Unsloth services
• Installation Fixes: Improved backend dependency installation when using local ART paths (#241)
• Documentation Updates: Various minor documentation improvements and clarifications

Technical Improvements

• Updated SkyPilot backend installation instructions
• Removed obsolete numpy installation cells from quickstart examples
• Enhanced dependency synchronization

v0.4.3

ART 0.4.3 Release Notes

Breaking Changes

SkyPilot is now an optional dependency. If you use SkyPilotBackend, you must now install ART with the skypilot extra:

# Before (no longer works)
pip install openpipe-art

# Now required for SkyPilotBackend users
pip install openpipe-art[skypilot]

What's Changed

Dependencies

• Moved SkyPilot dependencies (semver>=3.0.4 and skypilot==0.9.3) to an optional dependency group [skypilot] (#235)
• This reduces the default installation size for users who don't need SkyPilot functionality

Documentation Updates

• Updated installation instructions in all relevant documentation:
  • Installation + Setup guide
  • ART Backend documentation
  • Summarizer tutorial

Migration Guide

If you're using SkyPilotBackend in your code:

# Your existing code doesn't need to change, just update the installation
from art.skypilot import SkyPilotBackend
backend = SkyPilotBackend(...)

Simply install with: pip install openpipe-art[skypilot] or uv add openpipe-art[skypilot]

Full Changelog

See PR #235 for complete details.

v0.4.2

What's Changed

• Fix client import error by vendoring transformers constants (#232)
• docs: Add comprehensive documentation for additional_histories feature (#231)
• Fix Ruff lint (#229)
• Update 2048 code to use RULER (#228)
• Add RULER notebook for 2048 (#227)
• Add RULER promotional snippet to README (#225)
• Add run_checks.sh script for code quality checks (#224)
• Update README (#223)
• fix python version in art-e (#222)
• ruler docs (#221)
• feat: Decouple vLLM & Unsloth Trainer (#212)

Full Changelog: v0.4.0...v0.4.2

v0.4.1

What's Changed

• Fix client import error by vendoring transformers constants (#232)
• Fix Ruff lint (#229)
• Update 2048 code to use RULER (#228)
• Add RULER notebook for 2048 (#227)
• Add RULER promotional snippet to README (#225)
• Add run_checks.sh script for code quality checks (#224)
• Update README (#223)
• fix python version in art-e (#222)
• ruler docs (#221)
• feat: Decouple vLLM & Unsloth Trainer (#212)

Full Changelog: v0.4.0...v0.4.1

v0.4.0

🚀 Introducing RULER: Relative Universal LLM-Elicited Rewards

We're excited to announce ART v0.4.0, featuring RULER - a groundbreaking general-purpose reward function that makes agent training dramatically easier and faster!

📏 What is RULER?

RULER (Relative Universal LLM-Elicited Rewards) uses an LLM-as-judge to rank agent trajectories, eliminating the need for:

• ❌ Labeled training data
• ❌ Expert feedback
• ❌ Hand-crafted reward functions

Yet it often matches or exceeds the performance of carefully designed reward functions!

🎯 Key Benefits

• 2-3x faster development: Skip the tedious reward engineering phase
• Universal application: Works across diverse RL tasks without modification
• Production-ready: Battle-tested on real tasks with impressive results
• Simple integration: Just a few lines of code to get started

📖 Learn More

Check out the RULER documentation to see how easy it is to use:

from art.rewards import ruler_score_group

# Score your trajectories with one line
judged_group = await ruler_score_group(group, "openai/gpt-4o-mini")

Read the full launch announcement for detailed performance comparisons and insights.

What's Changed

Major Features

• Add RULER reward function (#218) 🎉
• RULER documentation (#221)

Other Improvements

• Update README (#223)
• fix python version in art-e (#222)
• Add setproctitle as dep in colab notebooks (#220)
• Move plotting dependencies to optional group (#217)
• feat: tau-bench brad 003 (#216)
• Allow validation_loader argument to train method (#215)
• Update swe-bench example docs (#214)
• chore: Remove workaround for torch-compile and use --torch-compile flag (#212)
• Adds option to use padding with --torch-compile (#211)
• Fix tau-bench example (#210)
• art-2048: update qwen model identifier (#209)
• Allow Unsloth to use --pad_token when tokenizer has no pad token (#208)
• Allow using specific wandb projects in the CLI (#207) (#207)
• feat: Allow using get_peft_model to re-initialize trainer state (#206)
• chore: Add art_trainer module with ART's TRL Trainer (#205)
• Art tau bench example (#204)
• 🔊 Improve noisy startup (#203) (#203)
• feat: SWE-Bench Example (#201)
• Update to 0.3.13, pin accelerate (#197)

Full Changelog: v0.3.13...v0.4.0
EOF < /dev/null

Release v0.3.13

What's Changed

• chore: Update TRL (#187)
• allow training without logprobs experimentation (#186)
• chore: Upgrade Unsloth dependencies (#183)
• chore: SWE-Bench related changes (#181)
• Bump uv to >=0.6.15 (#180)
• Tau bench async rl (#179)
• Track entropy at training time (#178)
• Adds vllm metrics to wandb (#177)
• Pin openpipe-art and accelerate versions in notebooks (#175)
• Release 0.3.12 (#174)
• Match default SkyPilotBackend version to client (#173)
• feat: Add support for multiple histories (#170)
• [WIP] More docs (#169)

Full Changelog: v0.3.12...v0.3.13

Release v0.3.12

What's Changed

• Tau bench async (#168)
• Refactor dev/tau-bench for true async (#167)
• ART-E updates (#166)
• Add langfuse tracing to run_rl.py (#165)
• Make rollout_tau_bench_task synchronous (#164)
• feat: Multi-device training (#163)
• Create run_training.py for remote training (#162)
• Create run_rl.py with ART RL loop (#161)
• Wandb weave (#158)
• Basic W&B Weave integration (#157)
• Properly read base model from CLI (#156)
• Deploy model locally (#155)
• Fix s3 utils typo (#153)
• Fix busy wait in vllm test client (#152)
• Fix comment (#151)
• dev: swebench (#149)
• fix: Improve retry util typing (#148)
• Add get_guided_completion_params and use in tic tac toe self play (#147)
• Pin vllm to 0.8.5 (#146)

Full Changelog: v0.3.11...v0.3.12

Release v0.3.11

What's Changed

• Limit number of metrics shown in gather_trajectory_groups (#145)

Full Changelog: v0.3.10...v0.3.11

Release v0.3.10

What's Changed

• Update package version to 0.3.9 (#143)
• Speed up step deployment (#142)
• Add tic tac toe self-play example (#141)
• Fix training stability issues with new vLLM version (#140)

Full Changelog: v0.3.9...v0.3.10

Release v0.3.9

What's Changed

• Serialize model config (#139)
• Properly log trajectories and metrics via remote backend (#138)
• Properly sync workdir (#137)
• reverting to older vllm version, since the latest one shows regressions in convergence of grpo training (#136)
• feat: Add force_restart option to SkypilotBackend (#135)
• Fix train step (#134)
• Revert some of the thread safe changes (#133)
• Enable asymmetric PPO clipping (#132)
• Support close method on remote backend (#129)
• Fix hanging (#126)
• update version (#122)
• Update to version 0.3.6 (#121)
• Simplify tic tac toe example (#120)
• add qwen 3 support (#118)

Full Changelog: v0.3.7...v0.3.9