unslothai/unsloth

How to set "reasoning_effort" of GPT-OSS during GRPO rollouts?

Open

#3949 opened on Jan 29, 2026

View on GitHub
 (3 comments) (0 reactions) (0 assignees)Python (64,271 stars) (5,658 forks)batch import
help wanted

Description

  1. Did you update? Yes
  2. Colab or Kaggle or local / cloud. Kaggle
  3. Number GPUs used, use nvidia-smi. 1
  4. Which trainer? GRPOTrainer

Question: We can see how to set reasoning_effort when manually inference in official colab example:

inputs = tokenizer.apply_chat_template(
    messages,
    add_generation_prompt = True,
    return_tensors = "pt",
    return_dict = True,
    reasoning_effort = "low", # **NEW!** Set reasoning effort to low, medium or high
).to("cuda")

_ = model.generate(**inputs, max_new_tokens = 64, streamer = TextStreamer(tokenizer))

But how to set reasoning_effort of GRPO Trainer (rollouts)? I could not find this option in official colab examples. I have tested that maybeGRPOTrainer is using "high" by default . But for me "medium" is enough and more time-efficient.

Thanks in advance!

Contributor guide

Tech stack
pythonpytorch
Domain
machine learning
Issue type
feature
DifficultyEstimated implementation difficulty for a new contributor, from 1 for very small changes to 5 for expert-level work.
2
Estimated timeA rough time range for an experienced contributor to investigate, implement, test, and prepare a pull request.
under 1 hour
Activity statusHow available the issue appears right now: fresh, active, stale, blocked, or waiting on maintainer input.
needs maintainer response
ClarityHow clearly the issue explains the expected change, acceptance criteria, and next step.
clear
Prerequisites
PythonPyTorchGRPO trainingunsloth
Newbie friendlinessA 1-100 score estimating how approachable this issue is for first-time contributors.
70
Research direction
Investigate the GRPOTrainer code in the unsloth repository. Check the generation config or the process function to see if reasoning effort is passed to the model. If not, a new parameter needs to be added to the trainer's init and propagate to the generate call. Also review any existing discussions in the issue comments for workarounds.