Create README.md
Browse files
README.md
ADDED
|
@@ -0,0 +1,35 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
license: apache-2.0
|
| 3 |
+
tags:
|
| 4 |
+
- reasoning
|
| 5 |
+
- mathematics
|
| 6 |
+
- reinforcement-learning
|
| 7 |
+
datasets:
|
| 8 |
+
- AIME
|
| 9 |
+
- AMC
|
| 10 |
+
- Omni-Math
|
| 11 |
+
base_model: R1-Distill-Qwen-1.5B
|
| 12 |
+
---
|
| 13 |
+
|
| 14 |
+
# ALP_R1_Qwen1.5B
|
| 15 |
+
|
| 16 |
+
R1-Distill-Qwen-1.5B trained with Adaptive Length Penalty (ALP) - reduces token usage by ~50% while maintaining performance.
|
| 17 |
+
|
| 18 |
+
## Training
|
| 19 |
+
- 100 steps GRPO, batch 512, LR 1e-6, β=1e-7
|
| 20 |
+
- 16 rollouts/prompt for difficulty estimation
|
| 21 |
+
- 8K context window
|
| 22 |
+
|
| 23 |
+
## Performance (Pass@1)
|
| 24 |
+
- MATH-500: 0.81
|
| 25 |
+
- AIME: 0.252
|
| 26 |
+
- OlympiadBench: 0.51
|
| 27 |
+
|
| 28 |
+
## Token Usage
|
| 29 |
+
- MATH: 2804→862 (-69%)
|
| 30 |
+
- AIME: 4007→3331 (-17%)
|
| 31 |
+
- Olympiad: 3606→2107 (-42%)
|
| 32 |
+
|
| 33 |
+
## Usage
|
| 34 |
+
```python
|
| 35 |
+
prompt = f"{problem} Let's think step by step and output the final answer within \\boxed{{}}."
|