---
license: apache-2.0
datasets:
- openai/gsm8k
language:
- en
base_model:
- Qwen/Qwen2.5-7B-Instruct
pipeline_tag: question-answering
tags:
- text-generation-inference
- open-llm-leaderboard
---
license: mit
base_model：

- Qwen/Qwen2.5-7B

---

参考simple_GROP项目训练的模型，GSM8K，训练了200个step，出现了一次however。
使用了3张A800 80G，训练了20多分钟

训练结果：

loss
<img src="https://huggingface.co/whysue/simple_GRPO/resolve/main/loss.png">

GPU
<img src="https://huggingface.co/whysue/simple_GRPO/resolve/main/GPU.png">

memory
<img src="https://huggingface.co/whysue/simple_GRPO/resolve/main/memory.png">

---

测试结果

demo_math_chat_gen(simple_GRPO_why)
<img src="https://huggingface.co/whysue/simple_GRPO/resolve/main/demo_math_chat_gen(simple_GRPO_why).png">
demo_math_chat_gen(Qwen2.5-7B)
<img src="https://huggingface.co/whysue/simple_GRPO/resolve/main/demo_math_chat_gen(Qwen2.5-7B).png">
notice

在GSM8K上进行评估，Qwen2.5-7B的得分为85.4。原因可能是是https://github.com/open-compass/opencompass/issues/1878

---