qwen2.5-32b-instruct-limo-lora-adapter

qwen2.5-32b-instruct-limo-lora-adapter

This model is a fine-tuned version of Qwen/Qwen2.5-32B-Instruct model. The fine-tuning was performed using Low-Rank Adaptation (LoRA) on the LIMO dataset to enhance the model's reasoning capabilities, based on the work in the paper: LIMO: Less is More for Reasoning.

Model description

• Base Model: Qwen/Qwen2.5-32B-Instruct
• Fine-Tuning Dataset: GAIR/LIMO
• Fine-Tuning Method: Low-Rank Adaptation (LoRA)
• Library Used: peft
• License: Apache 2.0

Usage

To utilize this model for text generation tasks, follow the steps below:

Installation

Ensure you have the necessary libraries installed:

pip install torch transformers peft

Generating Text

from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel
# Load the base model
base_model_name = "Qwen/Qwen2.5-32B-Instruct"
base_model = AutoModelForCausalLM.from_pretrained(base_model_name, torch_dtype="auto", device_map="auto")
# Load the tokenizer
tokenizer = AutoTokenizer.from_pretrained(base_model_name)
# Load the LoRA adapter
adapter_path = "t83714/qwen2.5-32b-instruct-limo-lora-adapter"
model = PeftModel.from_pretrained(base_model, adapter_path)
prompt = "How much is (2+5)x5/7"
# Tokenize the input
inputs = tokenizer(prompt, return_tensors="pt").to("cuda")
# Generate the output
output = model.generate(**inputs, max_length=8000)
print(tokenizer.decode(output[0], skip_special_tokens=True))

Merge the adapter and export merged model

from peft import PeftModel
from transformers import AutoModelForCausalLM
base_model = AutoModelForCausalLM.from_pretrained("Qwen/Qwen2.5-32B-Instruct")
# Load the LoRA adapter
adapter_path = "t83714/qwen2.5-32b-instruct-limo-lora-adapter"
model = PeftModel.from_pretrained(base_model, adapter_path)
merged_model = model.merge_and_unload()
merged_model.save_pretrained("./merged-model/")

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 5e-06
• train_batch_size: 1
• eval_batch_size: 8
• seed: 42
• distributed_type: multi-GPU
• optimizer: Use adamw_torch with betas=(0.9,0.999) and epsilon=1e-08 and optimizer_args=No additional optimizer arguments
• generation_max_length: 16384
• lr_scheduler_type: cosine
• num_epochs: 15
• lora rank: 8
• lora target layers:
  • v_proj
  • o_proj
  • q_proj
  • k_proj

Eval Result

Math 500 pass@1: 85%

Acknowledgment

This model is trained based on the work of Ye et al. (2025). If you use this model, please also consider citing their paper:

@misc{ye2025limoreasoning,
      title={LIMO: Less is More for Reasoning}, 
      author={Yixin Ye and Zhen Huang and Yang Xiao and Ethan Chern and Shijie Xia and Pengfei Liu},
      year={2025},
      eprint={2502.03387},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2502.03387}, 
}