TRL
Named by 18 kept items in this workspace.
16 items name it under frameworks and tools and 2 under classification.technologies.
Search every workspace for TRL. It matches the name as written, so it can find other things that share it.
The items that name it 18
Training library mentioned alongside Mistral Large 4 support.
Supports single-node LoRA SFT of Mistral Large 4 on 8×B300 at 142 GiB per GPU.
Hugging Face library used for day-0 supervised fine-tuning (SFTTrainer) of Mistral Large 4.
Hugging Face training library; async GRPO trainer now defaults to score centering.
Part of the PyTorch ecosystem named among libraries.
Hugging Face post-training library; SFTTrainer runs supervised fine-tuning on Trainium. From TRL 1.16, Neuron batches are padded to a multiple of 128 tokens.
SFTTrainer used for supervised fine-tuning of Beam with LoRA
Supervised fine-tuning of the full 816B checkpoint with LoRA.
Training library; v1.15 fused LM head on by default
Training library that records token IDs and logprobs from the capture proxy to turn harnesses into RL environments.
post-training library used as the throughput baseline Halo compares against
Hugging Face library providing AsyncGRPOTrainer, which now supports training a LoRA adapter and syncing only the adapter to vLLM (v1.14, via PR #7017)
Library used to run GRPO reinforcement learning training on the coding model, including the GRPOTrainer that was modified with four changes to make training work
Provides GRPOConfig/GRPO trainer used to run the reinforcement-learning fine-tuning of LFM2.5-350M on structured outputs
Hugging Face training library; Halo trainers can subclass TRL trainers, and is used as the stock baseline in benchmarks (trl.SFTTrainer).
Library providing preference trainers (DPO/ORPO/etc.) used by Soup's training stack
All technologies in this workspace.