ReasoningTransferability/UniReason-Qwen3-14B-think-SFT

UniReason-Qwen3-14B-think-SFT

This model is associated with the research paper: "Does Math Reasoning Improve General LLM Capabilities? Understanding Transferability of LLM Reasoning"

📄 Paper: 2507.00432 

Model Description

This model is a DISTILL FROM QWEN3-32B-INSTRUCT (THINKING MODE) THROUGH REJECT SAMPLING-tuned version of Qwen3-14B-Base focused on math-reasoning capabilities. The model was developed as part of research investigating the transferability of mathematical reasoning skills to general language tasks.

Key Research Questions Addressed:

• —Does math reasoning training improve general LLM capabilities?
• —How do different training methods (RL vs SFT) affect transferability?
• —What is the trade-off between specialized math performance and general capabilities?

Model Details

• —Base Model: Qwen3-14B-Base
• —Training Method: DISTILL FROM QWEN3-32B-INSTRUCT (THINKING MODE) THROUGH REJECT SAMPLING
• —Primary Focus: math-reasoning
• —Training Data: Math-specific datasets
• —Architecture: Transformer-based language model
• —Parameters: 14B

Training Details

Training Method: DISTILL FROM QWEN3-32B-INSTRUCT (THINKING MODE) THROUGH REJECT SAMPLING

Custom training methodology - see paper for details.

Datasets Used

• —Mathematical reasoning datasets
• —See paper for complete dataset list

Performance

Math Reasoning Benchmarks

• —MATH: See paper
• —AIME: See paper

General Capabilities

• —General QA: See paper
• —Code Generation: See paper
• —Instruction Following: See paper

For detailed performance metrics, please refer to the paper.

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

# Load model and tokenizer
model_name = "ReasoningTransferability/UniReason-Qwen3-14B-think-SFT"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForCausalLM.from_pretrained(
    model_name,
    torch_dtype=torch.float16,
    device_map="auto"
)

# Example: Math reasoning
math_prompt = "Solve this step by step: What is the derivative of x^3 + 2x^2 - 5x + 1?"
inputs = tokenizer(math_prompt, return_tensors="pt")
outputs = model.generate(**inputs, max_length=32768, temperature=0.7)
response = tokenizer.decode(outputs[0], skip_special_tokens=True)
print(response)

# Example: General reasoning
general_prompt = "Explain the concept of supply and demand in economics."
inputs = tokenizer(general_prompt, return_tensors="pt")
outputs = model.generate(**inputs, max_length=32768, temperature=0.7)
response = tokenizer.decode(outputs[0], skip_special_tokens=True)
print(response)

Limitations and Biases

• —Specialization Trade-offs: As explored in the paper, models optimized for math reasoning may show reduced performance on general tasks
• —Training Method Dependencies: Performance characteristics vary significantly between RL and SFT training approaches
• —Domain Transfer: The extent of capability transfer from math to other domains is limited
• —Computational Requirements: Model requires significant computational resources for inference

Research Findings

Key findings from the associated paper:

1. 1.RL vs SFT: RL-tuned models show better transfer to general domains compared to SFT-tuned models
2. 2.Capability Trade-offs: Most math-specialized models fail to transfer gains to other domains
3. 3.Forgetting: SFT-tuned models often forget general capabilities during math-focused training

Ethical Considerations

• —This model is intended for research purposes
• —Users should be aware of potential biases in mathematical and general reasoning
• —The model should not be used for making critical decisions without human oversight
• —Consider the environmental impact of large model inference

Contact

For questions about this model or the associated research, please:

• —Open an issue in this repository
• —Contact the paper authors
• —Reference the original paper: https://arxiv.org/abs/2507.00432

Acknowledgments

This work builds upon the research presented in "Does Math Reasoning Improve General LLM Capabilities? Understanding Transferability of LLM Reasoning" and uses the Qwen3-14B-Base architecture as its foundation.

Model uploaded on 2025-09-28