Adaptive Length Penalty
Collection
Teaching language models to think efficiently with Adaptive Length Penalty (ALP)
•
3 items
•
Updated
•
1
ALP_R1_Qwen1.5B
R1-Distill-Qwen-1.5B trained with Adaptive Length Penalty (ALP) - reduces token usage by ~50% while maintaining performance.
Training
- 100 steps GRPO, batch 512, LR 1e-6, β=1e-7
- 16 rollouts/prompt for difficulty estimation
- 8K context window
Performance (Pass@1)
- MATH-500: 0.81
- AIME: 0.252
- OlympiadBench: 0.51
Token Usage
- MATH: 2804→862 (-69%)
- AIME: 4007→3331 (-17%)
- Olympiad: 3606→2107 (-42%)
Usage
prompt = f"{problem} Let's think step by step and output the final answer within \\boxed{{}}."
- Downloads last month
- 260