---
license: apache-2.0
pipeline_tag: text-generation
---

# Model Card for Qwen3-32B-LoRA-ECHO-KK-GRPO

<!-- Provide a quick summary of what the model is/does. -->

Based on Qwen3-32B, we applied the ECHO framework to perform LoRA fine-tuning on the KK dataset. 
Ultimately, it achieved near-perfect scores on the 2–8 PPL test set, surpassing o4-mini, DeepSeek-R1, and o3-mini-high.

Tabel 3: Model performance on K&K logic puzzle task  across different degrees of difficulty
|  model    | 2 | 3 | 4 | 5 | 6 | 7 |  8 |
|----------------|---------------------------:|--------------------------:|--------------------------:|--------------------:|------------:|-------------:|-------------:|
| Qwen3-32B  | 0.98                      | 0.99                     | 0.98                     | 0.99               | 0.98        | 0.96         |0.95       |
| Deepseek-R1 | 1.00                   | 0.97                    | 0.95                    | 0.93              | 0.91       | 0.93         |0.91        |
| o3-mini-high  | 1.00                      | 1.00                     | 1.00                     | 1.00               | 0.99         | 0.98         |0.98         |
| o4-mini  | 1.00                      | 1.00                      | 0.96                     | 0.94               | 0.97        | 0.93         |0.87         |
| Qwen3-32B-Echo(GRPO w/Lora)  | 0.99                      | 1.00                     | 1.00                     | 1.00               | 0.99        | 1.00         |0.99         |

# Quick start



```python

from transformers import AutoModelForCausalLM, AutoTokenizer

model_name = "GradientResearch/Qwen3-32B-LoRA-ECHO-KK-GRPO"# load the tokenizer and the model
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForCausalLM.from_pretrained(
    model_name,
    torch_dtype="auto",
    device_map="auto"
)

# prepare the model input
prompt = "K & K"
messages = [
    {"role": "user", "content": prompt}
]
text = tokenizer.apply_chat_template(
    messages,
    tokenize=False,
    add_generation_prompt=True,
    enable_thinking=True # Switches between thinking and non-thinking modes. Default is True.
)
model_inputs = tokenizer([text], return_tensors="pt").to(model.device)

# conduct text completion
generated_ids = model.generate(
    **model_inputs,
    max_new_tokens=32768
)
output_ids = generated_ids[0][len(model_inputs.input_ids[0]):].tolist() 

# parsing thinking contenttry:
    # rindex finding 151668 (</think>)
    index = len(output_ids) - output_ids[::-1].index(151668)
except ValueError:
    index = 0

thinking_content = tokenizer.decode(output_ids[:index], skip_special_tokens=True).strip("\n")
content = tokenizer.decode(output_ids[index:], skip_special_tokens=True).strip("\n")

print("thinking content:", thinking_content)
print("content:", content)
```


# Citation
If you find our work helpful, feel free to give us a cite.
```
@misc{xiao2025echodecouplinginferencetraining,
      title={Echo: Decoupling Inference and Training for Large-Scale RL Alignment on Heterogeneous Swarms}, 
      author={Jie Xiao and Changyuan Fan and Qingnan Ren and Alfred Long and Yuchen Zhang and Rymon Yu and Eric Yang and Lynn Ai and Shaoduo Gan},
      year={2025},
      eprint={2508.05387},
      archivePrefix={arXiv},
      primaryClass={cs.LG},
      url={https://arxiv.org/abs/2508.05387}, 
}
```