trl
Train transformer language models with reinforcement learning.
File Explorer
Download Latest Version (.zip)- AGENTS.md
- BUGBOT.md
- custom-queries.qls
- bug-report.yml
- feature-request.yml
- new-trainer-addition.yml
- build_documentation.yml
- build_pr_documentation.yml
- clear_cache.yml
- codeQL.yml
- docker-build.yml
- pr_template_check.yml
- publish.yml
- slow-tests.yml
- sync-huggingface-skills.yml
- tests-experimental.yml
- tests.yml
- tests_latest.yml
- tests_python_versions.yml
- tests_transformers_branch.yml
- trufflehog.yml
- upload_pr_documentation.yml
- dependabot.yml
- PULL_REQUEST_TEMPLATE.md
- logo-dark.png
- logo-light.png
- Dockerfile
- Dockerfile
- _toctree.yml
- a2po_trainer.md
- async_distillation_trainer.md
- async_grpo_trainer.md
- bco_trainer.md
- bema_for_reference_model.md
- callbacks.md
- chat_template_utils.md
- chat_templates.md
- clis.md
- community_tutorials.md
- cpo_trainer.md
- customization.md
- data_utils.md
- dataset_formats.md
- deepspeed_integration.md
- distillation_trainer.md
- distributing_training.md
- dpo_trainer.md
- example_overview.md
- experimental_overview.md
- gkd_trainer.md
- gmpo.md
- gold_trainer.md
- grpo_trainer.md
- grpo_with_replay_buffer.md
- gspo_token.md
- harbor.md
- index.md
- installation.md
- jobs_training.md
- kernels_hub.md
- kto_trainer.md
- liger_kernel_integration.md
- lora_without_regret.md
- merge_model_callback.md
- minillm_trainer.md
- nash_md_trainer.md
- online_dpo_trainer.md
- openenv.md
- openreward.md
- orpo_trainer.md
- paper_index.md
- peft_integration.md
- ppo_trainer.md
- prm_trainer.md
- quickstart.md
- rapidfire_integration.md
- reducing_memory_usage.md
- reward_trainer.md
- rewards.md
- rloo_trainer.md
- script_utils.md
- sdft_trainer.md
- sdpo_trainer.md
- sft_trainer.md
- speeding_up_training.md
- ssd_trainer.md
- tpo_trainer.md
- trackio_integration.md
- unsloth_integration.md
- usage_stats.md
- use_model.md
- vllm_integration.md
- xpo_trainer.md
- alst_ulysses_4gpu.yaml
- context_parallel_2gpu.yaml
- deepspeed_zero1.yaml
- deepspeed_zero2.yaml
- deepspeed_zero3.yaml
- fsdp1.yaml
- fsdp2.yaml
- multi_gpu.yaml
- single_gpu.yaml
- example_config.yaml
- deepmath_103k.py
- hh-rlhf-helpful-base.py
- llava_instruct_mix.py
- lm-human-preferences-descriptiveness.py
- lm-human-preferences-sentiment.py
- math_shepherd.py
- prm800k.py
- rlaif-v.py
- tldr.py
- tldr_preference.py
- ultrafeedback-prompt.py
- ultrafeedback.py
- grpo_agent.ipynb
- grpo_ministral3_vl.ipynb
- grpo_qwen3_vl.ipynb
- grpo_rnj_1_instruct.ipynb
- grpo_trl_lora_qlora.ipynb
- openenv_sudoku_grpo.ipynb
- openenv_wordle_grpo.ipynb
- README.md
- sft_ministral3_vl.ipynb
- sft_nemotron_3.ipynb
- sft_qwen_vl.ipynb
- sft_tool_calling.ipynb
- sft_trl_lora_qlora.ipynb
- __init__.py
- README.md
- __init__.py
- env.py
- kernel_server.py
- README.md
- run_cell.py
- __init__.py
- env.py
- README.md
- __init__.py
- data_agent.py
- browsergym.py
- browsergym_llm.py
- carla.py
- carla_vlm.py
- carla_vlm_gemma.py
- catch.py
- echo.py
- multi_env.py
- opencode.py
- opencode_hf_sandbox.py
- sudoku.py
- sudoku_prompt.txt
- wordle.py
- seta.py
- ppo.py
- ppo_tldr.py
- async_distillation.py
- async_distillation_mopd.py
- async_grpo.py
- bco.py
- cpo.py
- distillation.py
- dpo.py
- dpo_vlm.py
- gkd.py
- gold.py
- grpo_2048.py
- grpo_agent.py
- grpo_continuous_batching.py
- grpo_vlm.py
- gspo.py
- gspo_vlm.py
- kto.py
- mpo_vlm.py
- nash_md.py
- online_dpo.py
- online_dpo_vlm.py
- orpo.py
- prm.py
- reward_modeling.py
- rloo.py
- rloo_vlm.py
- sdft.py
- sdpo.py
- sft.py
- sft_diffusion_gemma.py
- sft_gemma3.py
- sft_gpt_oss.py
- sft_nemotron_3.py
- sft_tiny_aya_tool_calling.py
- sft_vlm.py
- sft_vlm_gemma3.py
- ssd.py
- ssd_eval.py
- tiny_aya_chat_template.jinja
- tpo.py
- xpo.py
- README.md
- __init__.py
- diffusion_gemma_for_block_diffusion.py
- __init__.py
- configuration_remote.py
- modeling_remote.py
- __init__.py
- cohere2_for_causal_lm.py
- cohere_for_causal_lm.py
- deepseek_v3_for_causal_lm.py
- deepseek_v3_for_causal_lm_0528.py
- falcon_mamba_for_causal_lm.py
- gemma2_for_causal_lm.py
- gemma_for_causal_lm.py
- glm4_moe_for_causal_lm.py
- gpt2_lm_head_model.py
- gpt_neox_for_causal_lm.py
- gpt_oss_for_causal_lm.py
- lfm2_for_causal_lm.py
- lfm2_for_causal_lm_2_5.py
- llama_for_causal_lm_3.py
- llama_for_causal_lm_3_1.py
- llama_for_causal_lm_3_2.py
- mistral_for_causal_lm_0_1.py
- mistral_for_causal_lm_0_2.py
- nemotron_3_5_lightning_for_causal_lm.py
- nemotron_3_nano_for_causal_lm.py
- nemotron_3_super_for_causal_lm.py
- nemotron_3_ultra_for_causal_lm.py
- olmo3_for_causal_lm.py
- opt_for_causal_lm.py
- peft_qwen3_for_causal_lm.py
- peft_qwen3_for_causal_lm_2.py
- phi3_for_causal_lm_3.py
- phi3_for_causal_lm_3_5.py
- qwen2_for_causal_lm_2_5.py
- qwen2_for_causal_lm_2_5_coder.py
- qwen2_for_causal_lm_r1_distill.py
- qwen3_for_causal_lm.py
- qwen3_for_causal_lm_instruct_2507.py
- qwen3_moe_for_causal_lm.py
- qwen3_moe_for_causal_lm_response_template.py
- remote_for_causal_lm.py
- small_qwen2_for_causal_lm_2_5.py
- small_qwen3_for_causal_lm.py
- __init__.py
- bart_model.py
- gemma3_for_conditional_generation.py
- gemma4_for_conditional_generation.py
- idefics2_for_conditional_generation.py
- idefics3_for_conditional_generation.py
- internvl_for_conditional_generation.py
- lfm2_vl_for_conditional_generation.py
- llava_for_conditional_generation.py
- llava_next_for_conditional_generation.py
- muse_glimmer_for_conditional_generation.py
- paligemma_for_conditional_generation.py
- qwen2_5_vl_for_conditional_generation.py
- qwen2_vl_for_conditional_generation.py
- qwen3_5_for_conditional_generation_nothink.py
- qwen3_5_for_conditional_generation_think.py
- qwen3_6_for_conditional_generation.py
- qwen3_8_for_conditional_generation.py
- qwen3_vl_for_conditional_generation.py
- smolvlm_for_conditional_generation.py
- t5_for_conditional_generation.py
- __init__.py
- gpt_neox_for_sequence_classification.py
- llama_for_sequence_classification_3_2.py
- qwen2_for_sequence_classification_2_5.py
- qwen3_for_sequence_classification.py
- qwen3_moe_for_sequence_classification.py
- __init__.py
- _common.py
- README.md
- add_copyrights.py
- generate_harmony_dataset.py
- generate_toolcall_dataset.py
- generate_zen_dataset.py
- generate_zen_image_dataset.py
- generate_zen_multi_image_dataset.py
- log_reports.py
- template.jinja
- ddp.yaml
- fsdp2.yaml
- zero2.yaml
- zero3.yaml
- __init__.py
- test_distributed.py
- __init__.py
- _openreward_echo_env.py
- test_a2po_trainer.py
- test_async_distillation_trainer.py
- test_async_grpo_trainer.py
- test_bco_trainer.py
- test_cpo_trainer.py
- test_gkd_trainer.py
- test_gmpo_trainer.py
- test_gold_trainer.py
- test_grpo_with_replay_buffer_trainer.py
- test_gspo_token_trainer.py
- test_harbor.py
- test_iw_opd_trainer.py
- test_merge_model_callback.py
- test_minillm_trainer.py
- test_modeling_value_head.py
- test_nash_md_trainer.py
- test_online_dpo_trainer.py
- test_openreward.py
- test_orpo_trainer.py
- test_ppo_trainer.py
- test_prm_trainer.py
- test_sdft_trainer.py
- test_sdpo_trainer.py
- test_self_distillation_trainer_behavior.py
- test_server_distillation_trainer.py
- test_ssd_trainer.py
- test_tpo_trainer.py
- test_utils.py
- test_xpo_trainer.py
- dpo.json
- sft.json
- sft_fa2.json
- __init__.py
- README.md
- test_invariant.py
- __init__.py
- conftest.py
- test_activation_offloading.py
- test_callbacks.py
- test_chat_template_utils.py
- test_cli.py
- test_cli_utils.py
- test_data_utils.py
- test_distillation_trainer.py
- test_dpo_trainer.py
- test_grpo_trainer.py
- test_kto_trainer.py
- test_model_utils.py
- test_reward_trainer.py
- test_rewards.py
- test_rich_progress_callback.py
- test_rloo_trainer.py
- test_sft_trainer.py
- test_skills.py
- test_skills_cli.py
- test_utils.py
- test_vllm_client_server.py
- testing_constants.py
- testing_utils.py
- fsdp1.yaml
- fsdp2.yaml
- multi_gpu.yaml
- single_gpu.yaml
- zero1.yaml
- zero2.yaml
- zero3.yaml
- cohere.jinja
- cohere2.jinja
- cohere2_training.jinja
- cohere_training.jinja
- deepseek_r1_distill.jinja
- deepseek_r1_distill_training.jinja
- deepseekv3.jinja
- deepseekv3_training.jinja
- diffusion_gemma.jinja
- diffusion_gemma_training.jinja
- gemma.jinja
- gemma3.jinja
- gemma3_training.jinja
- gemma4.jinja
- gemma_training.jinja
- glm4moe.jinja
- glm4moe_training.jinja
- gptoss.jinja
- gptoss_training.jinja
- idefics3.jinja
- idefics3_training.jinja
- lfm2.jinja
- lfm2_2_5.jinja
- lfm2_2_5_vl.jinja
- lfm2_training.jinja
- llama3.jinja
- llama3_1.jinja
- llama3_2.jinja
- llama3_training.jinja
- llava_next.jinja
- llava_next_training.jinja
- muse_glimmer.jinja
- muse_glimmer_training.jinja
- nemotron_3_5_lightning.jinja
- nemotron_3_5_lightning_training.jinja
- nemotron_3_nano.jinja
- nemotron_3_nano_training.jinja
- nemotron_3_super.jinja
- nemotron_3_super_training.jinja
- nemotron_3_ultra.jinja
- nemotron_3_ultra_training.jinja
- phi3.jinja
- phi3_5.jinja
- phi3_5_training.jinja
- phi3_training.jinja
- qwen2_5.jinja
- qwen2_5_training.jinja
- qwen2_5_vl.jinja
- qwen2_5_vl_training.jinja
- qwen3.jinja
- qwen3_5_nothink.jinja
- qwen3_5_nothink_training.jinja
- qwen3_5_think.jinja
- qwen3_5_think_training.jinja
- qwen3_6.jinja
- qwen3_6_training.jinja
- qwen3_8.jinja
- qwen3_8_training.jinja
- qwen3_instruct_2507.jinja
- qwen3_instruct_2507_training.jinja
- qwen3_training.jinja
- qwen3_vl.jinja
- qwen3_vl_training.jinja
- README.md
- __init__.py
- base.py
- env.py
- skills.py
- training.py
- vllm_serve.py
- __init__.py
- accelerate_config.py
- accelerate_launcher.py
- main.py
- __init__.py
- a2po_config.py
- a2po_trainer.py
- __init__.py
- async_distillation_config.py
- async_distillation_trainer.py
- async_rollout_worker.py
- vllm_client.py
- weight_transfer.py
- __init__.py
- async_grpo_config.py
- async_grpo_trainer.py
- async_rollout_worker.py
- openenv_harness.py
- vllm_client.py
- weight_transfer.py
- __init__.py
- bco_config.py
- bco_trainer.py
- __init__.py
- callback.py
- dpo_trainer.py
- __init__.py
- cpo_config.py
- cpo_trainer.py
- __init__.py
- __init__.py
- gkd_config.py
- gkd_trainer.py
- __init__.py
- gmpo_config.py
- gmpo_trainer.py
- __init__.py
- gold_config.py
- gold_trainer.py
- gold_vlm.py
- __init__.py
- grpo_with_replay_buffer_config.py
- grpo_with_replay_buffer_trainer.py
- __init__.py
- grpo_trainer.py
- __init__.py
- _env.py
- _spec.py
- __init__.py
- iw_opd_config.py
- iw_opd_trainer.py
- __init__.py
- __init__.py
- minillm_config.py
- minillm_trainer.py
- __init__.py
- nash_md_config.py
- nash_md_trainer.py
- __init__.py
- online_dpo_config.py
- online_dpo_trainer.py
- __init__.py
- _spec.py
- environment.py
- __init__.py
- orpo_config.py
- orpo_trainer.py
- __init__.py
- modeling_value_head.py
- ppo_config.py
- ppo_trainer.py
- __init__.py
- prm_config.py
- prm_trainer.py
- __init__.py
- loss_utils.py
- sdft_config.py
- sdft_trainer.py
- teacher_sync.py
- __init__.py
- loss_utils.py
- sdpo_config.py
- sdpo_trainer.py
- teacher_sync.py
- __init__.py
- server_distillation_config.py
- server_distillation_trainer.py
- __init__.py
- ssd_config.py
- ssd_trainer.py
- __init__.py
- tpo_config.py
- tpo_trainer.py
- __init__.py
- xpo_config.py
- xpo_trainer.py
- __init__.py
- merge_model_callback.py
- utils.py
- __init__.py
- dataset_formatting.py
- profiling.py
- __init__.py
- vllm_client.py
- vllm_generation.py
- __init__.py
- activation_offloading.py
- utils.py
- __init__.py
- accuracy_rewards.py
- format_rewards.py
- other_rewards.py
- __init__.py
- _hf_argparser.py
- distillation.py
- dpo.py
- env.py
- grpo.py
- kto.py
- reward.py
- rloo.py
- sft.py
- utils.py
- vllm_serve.py
- SKILL.md
- __init__.py
- cli.py
- skills.py
- completions_dataset_card.md
- lm_model_card.md
- rm_model_card.md
- __init__.py
- base_config.py
- base_trainer.py
- callbacks.py
- distillation_config.py
- distillation_trainer.py
- dpo_config.py
- dpo_trainer.py
- grpo_config.py
- grpo_trainer.py
- kto_config.py
- kto_trainer.py
- model_config.py
- reward_config.py
- reward_trainer.py
- rloo_config.py
- rloo_trainer.py
- sft_config.py
- sft_trainer.py
- utils.py
- __init__.py
- _compat.py
- _lazy_module.py
- chat_template_utils.py
- data_utils.py
- distributed.py
- import_utils.py
- py.typed
- .gitignore
- .pre-commit-config.yaml
- AGENTS.md
- CITATION.cff
- CLAUDE.md
- CODE_OF_CONDUCT.md
- CONTRIBUTING.md
- LICENSE
- Makefile
- MANIFEST.in
- MIGRATION.md
- pyproject.toml
- README.md
- RELEASE.md
- SECURITY.md
- VERSION
# Installation Guide
1. Get the code
git clone https://github.com/huggingface/trl
Downloads the entire project code from GitHub to your computer.
cd trl
Moves into the project folder you just downloaded.
2. Official Install Script
Easy RecommendedPrerequisites
- Python 3 Python is required to use pip.
pip install trl
Installs the package published on PyPI directly โ no need to clone the source.
pip install git+https://github.com/huggingface/trl.git
Installs the package published on PyPI directly โ no need to clone the source.
After installing, open a new terminal and run the program's version command (e.g. --version) to confirm it worked.
Pulled directly from this repo's README.
3. Docker
EasyPrerequisites
- Git Needed to download the project code from GitHub.
- Docker Desktop Needed to build and run containers. Install it and keep it running in the background.
โ ๏ธ This is a large repository, so this method may point to an internal sub-package rather than the actual core product. Check the full README as well.
docker build -f docker/trl-dev/Dockerfile -t trl .
Builds a runnable image based on the Dockerfile.
docker run -p 8080:80 trl
Runs the built image as an actual container.
Run docker compose ps to check the containers are Up. If the README mentions a port, open http://localhost:PORT in your browser.
4. Python
EasyPrerequisites
pip install trl
Installs the package published on PyPI directly โ no need to clone the source.
pip install git+https://github.com/huggingface/trl.git
Installs the package published on PyPI directly โ no need to clone the source.
pip install -e .[dev]
Installs the Python libraries listed in requirements.txt (or similar).
If it runs without errors and prints output in the terminal, it worked.
Pulled directly from this repo's README.
5. Make
MediumPrerequisites
- Git Needed to download the project code from GitHub.
- Make Usually pre-installed on Linux/macOS. On Windows, install separately (e.g. via MSYS2 or WSL).
make
Compiles the code based on the generated build configuration to produce an executable.
If it finishes without errors, it worked. Try running the generated executable directly.
// repository documentation
Was this content helpful?
(0 ratings)