Search results

CI fails with test dependencies: AssertionError: Parameter has not changed

Status: Open.

Bug

#4748 In huggingface/trl;

· albertvillanova opened on Dec 24, 2025

Addition of Supervised Reinforcement Learning

Status: Open.

#4747 In huggingface/trl;

· kushalgarg101 opened on Dec 24, 2025

examples/notebooks/grpo_qwen3_vl.ipynb fails with TypeError: string indices must be integers, not 'str'

Status: Open.

#4746 In huggingface/trl;

· ympan0508 opened on Dec 24, 2025

completion_only_loss=True give wrong training steps

Status: Open.

#4738 In huggingface/trl;

· HuangChiEn opened on Dec 23, 2025

NCCL error: /pytorch/torch/csrc/distributed/c10d/NCCLUtils.hpp:268, NCCL version 2.22.3 ncclSystemError: System call (e.g. socket, malloc) or external library call failed or device error. Last error: socketStartConnect: Connect to IP<51423> failed

Status: Open.

#4737 In huggingface/trl;

· guotong1988 opened on Dec 23, 2025

feat GRPOTrainer: Bidirectional MIS from IcePop loss (used in Xiaomi MiMo-V2-Flash and Ring-1T)

Status: Open.

#4735 In huggingface/trl;

· casinca opened on Dec 22, 2025

change in behavior of mean_token_accuracy from trl==0.15 to trl==0.26 when using liger-kernel

⚡ PEFTRelated to PEFTRelated to PEFT

⚡accelerateRelated to accelerateRelated to accelerate

🐛 bugSomething isn't workingSomething isn't working

Status: Open.

#4730 In huggingface/trl;

· mvish7 opened on Dec 20, 2025

gkd cannot run on the dataset OpenThoughts3-1.2M.

⚡accelerateRelated to accelerateRelated to accelerate

🏋 GKDRelated to GKDRelated to GKD

🐛 bugSomething isn't workingSomething isn't working

Status: Open.

#4725 In huggingface/trl;

· Jim2016713 opened on Dec 19, 2025

Loss calculation of <code>GKDTrainer</code> may be inaccurate when performing gradient accumulation?

🏋 GKDRelated to GKDRelated to GKD

🐛 bugSomething isn't workingSomething isn't working

Status: Open.

#4719 In huggingface/trl;

· jue-jue-zi opened on Dec 18, 2025

Add RLVR (Reinforcement Learning with Verifiable Rewards) Trainer

✨ enhancementNew feature or requestNew feature or request

🏋 RewardRelated to Reward modellingRelated to Reward modelling

🏋 RLOORelated to RLOORelated to RLOO

Status: Open.

#4711 In huggingface/trl;

· mmjerge opened on Dec 17, 2025

GRPOTrainer: Relative LLM-as-a-judge rankings

✨ enhancementNew feature or requestNew feature or request

🏋 GRPORelated to GRPORelated to GRPO

Status: Open.

#4707 In huggingface/trl;

· austinmw opened on Dec 16, 2025

feat GRPOTrainer: Off-policy sequence masking from DeepSeek V3.2

✨ enhancementNew feature or requestNew feature or request

🏋 GRPORelated to GRPORelated to GRPO

Status: Open.

#4697 In huggingface/trl;

· casinca opened on Dec 15, 2025