Method
Online GRPO from the base model with a group relative length reward. Correct answers earn 1 plus a bonus that grows as they get shorter than their siblings; wrong answers never earn a positive reward. Uses the Dr. GRPO loss with no reward scaling, so long wrong answers are not favored.
Training data: AmareshHebbar/ShortThink_Traces_Qwen3_1.7B. Weights: LoRA on all linear layers, merged into full bf16 safetensors.
Why finetune instead of prompting
A prompt can ask for brevity, and the base model can switch thinking off entirely. The table below includes both of those, plus budget forcing (cap the thinking at 1024 tokens and force an answer), so you can see exactly what training buys over prompting.
Results
All rows use the same prompts, sampling settings and math_verify grading. MATH500 and GSM8K use 1 samples per problem, AIME 2024 uses 4. Tokens are generated tokens including the thinking. Tokens per correct is total tokens divided by correct answers, averaged over MATH500 and GSM8K. Latency is measured at batch size 1 with vLLM; cost assumes $1.0/hour GPU at the measured batched throughput.

Compression by difficulty (MATH500)

Training hyperparameters
Usage
from vllm import LLM, SamplingParams llm = LLM("AmareshHebbar/ShortThink_Qwen3_1.7B_GRPO")tok = llm.get_tokenizer()question = "What is 15% of 80?"msgs = [{"role": "user", "content": question + "\nPlease reason step by step, and put your final answer within \\boxed{}."}]prompt = tok.apply_chat_template(msgs, tokenize=False, add_generation_prompt=True, enable_thinking=True)out = llm.generate([prompt], SamplingParams(temperature=0.6, top_p=0.95, top_k=20, max_tokens=8192))print(out[0].outputs[0].text)
Use the same prompt format and sampling settings as the base model's thinking mode.
Files
The repository root holds merged bf16 safetensors. adapter/ holds the LoRA adapter where one was trained, training/ holds loss curves and logs, and eval/ holds raw evaluation results and sample generations.