Code: grpo-tldr.py

This LoRA aims to make the length of the TL;DR more reliable, targetting 10 words, and between 30 and 60 characters.

You can find the weight on HuggingFace

This is a almost a copy/paste of HuggingFace GRPO example. 1100 steps are enough for a usable LoRA, which took 1hr40min on my RTX3090 for Qwen3-0.6B.

Usage:

<|im_start|>user

Your role is to make a summary of the following post.

insert <post here>

TL;DR:
<|im_end|>
<|im_start|>assistant
<think>

</think>