The purpose of this issue is to list the tasks that need to be completed before we reach v1. This list is evolving and is modified based on recent discussions and progress.

Documentation

• Remove how_to_train.md Remove how_to_train.md: outdated training FAQ #4267
• Remove using_llama_models.md Remove using_llama_models.md: outdated Llama2-specific documentation #4268
• Remove logging.md Remove logging.md: trainer-specific metrics documentation #4269
• Rewrite peft_integration.md #4376
• Remove guidance about converting conversational to standard. #4375
• Move every section of conceptual_guides/experimental into its own section in experimental #4377
• Extend basic usage example to all supported CLIs #4378
• Remove or populate "Training customization" #4379
• Remove outdated warning about batch contamination #4381
• Populate "Speeding Up Training" #4382
• Add PEFT subsection to "Reducing Memory Usage" #4383
• Write the subsection "Multi-Node Training" #4384
• Use a common 'trl-lib` namespace for the models/datasets/spaces #4385
• Reference supported trainers in Liger Kernel integration guide #4386
• Remove Sentiment Tuning Examples #4396
• Remove or move Multi Adapter RL #4397
• Complete paper index #4407

Examples

• Update or remove some of the notebooks #4399
• Remove setting chat template in most trainers #4404

Tests

• Move everything in test/slow to their respective trainer test files #4401

Main codebase

• Add accuracy reward to the trl.rewards module Add accuracy reward #4270
• Add an option (default to True) to use RichProgressCallback in scripts (trl.scripts).
• Add kernels to Docker images #4398
• Remove log_example_reports.py Remove unused log_example_reports.py script #4241
• Remove commands directory Remove unused commands directory #4258
• Remove examples/research_projects Remove unused commands directory #4258
• Remove trl.extra.dataset_formatting Deprecate unused dataset_formatting module #4242 Remove deprecated functions and parameters #4651
• Remove support for FSDP1 #4387 (we may drop the support for FSDP1 post v1)
• Remove BestOfNSampler. Deprecate BestOfNSampler #4291 ⚰️ Remove deprecated #4301
• Fully transition from flash-attn to kernels #4380
• Move masked_mean, masked_var and masked_whiten to ppo.py #4403
• Refactor DPO to align implementation with SFT (WIP in Refactor DPO #3906)
• Tool calling for GRPO/RLOO (WIP in 🕵️‍♂️ GRPO: Agent training #4300)
• Async generation for Online methods
• Bump transformers to v5 we will keep v4 compatibility
• Make vLLM server OpenAI-compatible #4402 (we may do this post v1)

Moving experimental features to experimental submodule

Discussed in #4223 for trainers

• Move BCO to experimental submodule 🚚 Move BCO to trl.experimental #4312
• Move KTO to experimental submodule 🚚 Move KTO to trl.experimental #4575
• Move Nash-MD to experimental submodule Move NashMDTrainer to experimental module #4477
• Move ORPO to experimental submodule [ORPO] Move ORPOTrainer to experimental #4480
• Move Online DPO to experimental submodule #4472
• Move PPO to experimental submodule Move PPOTrainer to trl.experimental.ppo #4482
• Move PRM to experimental submodule Move PRMTrainer to trl.experimental.prm #4483
• Move XPO to experimental submodule Move XPOTrainer to trl.experimental.xpo #4485
• Move everything related to Mergekit to the experimental submodule #4395
• Move everything related to Judges to trl.experimental #4400
• Move Winrate callback to experimental Move WinRateCallback to experimental #4558