divelab/DAPO_E2H-countdown-gaussian_0p5_0p5

Overview

This model, divelab/DAPO_E2H-countdown-gaussian_0p5_0p5, is a specialized 1.5 billion parameter instruction-tuned language model. It is a fine-tuned variant of the Qwen/Qwen2.5-1.5B-Instruct base model, developed by divelab.

Key Capabilities

• Mathematical Reasoning: The model is specifically fine-tuned on the Countdown dataset, enhancing its capabilities in mathematical problem-solving and reasoning tasks.
• E2H Framework: Training utilized the E2H (Easy to Hard Reasoning) framework, which is designed to improve LLM reasoning through curriculum reinforcement learning.
• GRPO Method: It incorporates the GRPO (Gaussian Reward Policy Optimization) method, as introduced in the DeepSeekMath paper, further optimizing its mathematical reasoning performance.

Good For

• Mathematical Problem Solving: Ideal for applications requiring robust mathematical reasoning, such as solving arithmetic problems, logical puzzles, or tasks from the Countdown dataset.
• Research in Reasoning: Useful for researchers exploring advanced reasoning techniques in language models, particularly those interested in curriculum learning and reinforcement learning from human feedback (RLHF) applied to mathematical domains.