From e34066745ff06a11a4f6ae110391daf9a2e18326 Mon Sep 17 00:00:00 2001 From: sergiopaniego Date: Thu, 15 Oct 2026 17:46:30 +0200 Subject: [PATCH] Fix the default IS mode in the GRPOConfig docstring --- trl/trainer/grpo_config.py | 3 +-- 1 file changed, 1 insertion(+), 2 deletions(-) diff --git a/trl/trainer/grpo_config.py b/trl/trainer/grpo_config.py index a1f102386aa..fd81e7120c4 100644 --- a/trl/trainer/grpo_config.py +++ b/trl/trainer/grpo_config.py @@ -328,8 +328,7 @@ class GRPOConfig(_BaseConfig): Specifies how Importance Sampling is performed when `vllm_importance_sampling_correction=True`. Possible values are: - - `"token_truncate"`: Token-level truncated IS (default). Per-token ratios are clipped to - [C_min, C_max]. + - `"token_truncate"`: Token-level truncated IS. Per-token ratios are clipped to [C_min, C_max]. - `"token_mask"`: Token-level masked IS. Per-token ratios outside [C_min, C_max] are set to zero. - `"sequence_truncate"`: Sequence-level truncated IS. A single sequence ratio is clipped to [C_min, C_max] and applied to all tokens in the sequence.