A reasoning-enhanced language model fine-tuned from Qwen2.5-7B-Instruct using GRPO (Group Relative Policy Optimization) for special education applications.

Model Description

This model is trained to provide direct, concise answers without explicit chain-of-thought reasoning steps (NoThink variant). It focuses on generating accurate responses efficiently.

Base Model: Qwen/Qwen2.5-7B-Instruct
Training Method: GRPO (Group Relative Policy Optimization)
Training Steps: 300
Focus: Direct answer generation without verbose reasoning

Training Details

Training Configuration

Framework: veRL (Volcano Engine Reinforcement Learning)
Algorithm: GRPO
Batch Size: Configured for 4x GPU setup
Precision: bfloat16

Training Data

Educational reasoning tasks
Mathematical problem solving
General knowledge Q&A

Usage

python

from transformers import AutoModelForCausalLM, AutoTokenizer

model_name = "OpenLearnLM/special-r1-qwen2.5-7b-nothink"

tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForCausalLM.from_pretrained(
    model_name,
    torch_dtype="auto",
    device_map="auto"
)

messages = [
    {"role": "user", "content": "What is the capital of France?"}
]

text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tokenizer(text, return_tensors="pt").to(model.device)

outputs = model.generate(**inputs, max_new_tokens=512)
response = tokenizer.decode(outputs[0][inputs.input_ids.shape[1]:], skip_special_tokens=True)
print(response)

Model Variants

Model	Description
special-r1-qwen2.5-7b-nothink (this)	Direct answers without explicit reasoning
special-r1-qwen2.5-7b-think	With chain-of-thought reasoning

Limitations

Trained primarily on English and Korean data
May not perform optimally on highly specialized domains outside training distribution
As an early checkpoint (step 300), performance may improve with continued training

Citation

If you use this model, please cite:

bibtex

@misc{openlearnlm2025special,
  title={Special-R1: Reasoning Models for Education},
  author={OpenLearnLM Team},
  year={2025},
  publisher={HuggingFace},
  url={https://huggingface.co/OpenLearnLM/special-r1-qwen2.5-7b-nothink}
}

License

This model is released under the Apache 2.0 License, following the base model's license.

OpenLearnLM/special-r1-deepseek-qwen3-8b-sped-adaptive-think-reward

Special-R1-Qwen2.5-7B-NoThink