Qwen2.5-3B-Instruct-rt-2708

Model Details

Qwen2.5-3B-Instruct-rt-2708 is a 3B parameter language model designed for English text generation tasks. This model builds upon Qwen/Qwen2.5-3B-Instruct and is further fine-tuned on the jjzha/rt-2708 dataset. It focuses on enhancing factual reasoning abilities in generated text.

Model Developers

This model was fine-tuned by independent contributors using the Hugging Face Transformers library.

Variations

This is a fine-tuned version of the Qwen2.5-3B-Instruct model. No additional variants or intermediate checkpoints are currently provided.

Input

Text only.

Output

Text only.

Model Architecture

The model is an auto-regressive, transformer-based language model, fine-tuned with supervised learning to improve instruction-following and reasoning capabilities in English.

Model Dates

Fine-tuning was performed in February-April 2025. The base and instruct model was originally released by the Qwen team.

License

This model is released under the Apache 2.0 license.

Research Paper

Scaling Reasoning can Improve Factuality in Large Language Models

Intended Use & Limitations

Intended Use Cases

This model is intended for English language text generation tasks that require improved factual accuracy and reasoning. It is suitable for research, experimentation, and development of assistant-like chat applications.

The instruction-tuned base model follows the Qwen instruction format, and this fine-tuned version preserves that behavior.

Limitations

Despite improvements, the model may still produce factually incorrect or logically inconsistent outputs. It is not recommended for high-stakes decision-making applications without human oversight. Always verify generated content before relying on it in critical scenarios.

Hardware and Software

Training Factors

Fine-tuning was performed using the Hugging Face Transformers library and Pytorch FSDP. We used a multinode and multigpu setup with AMD MI250x GPUs.

Carbon Footprint

We only have aggregated statistics of all models fine-tuned and inferences. A cumulative of 6,500 GPU hours of computation was performed on AMD MI250x GPU modules, which has a TDP of 500 Watts. The experiments were ran from February to April 2025. During this time, the average carbon efficiency in Finland was 0.085 kg/kW h. This means we released about 276 kg of CO2 equivalent.

Training Data

Overview

Fine-tuning was performed on the jjzha/rt-2708 dataset, which focuses on enhancing reasoning and factual accuracy.

Evaluation Results

See paper for results.

Citation

@misc{zhang2025scalingreasoningimprovefactuality,
      title={Scaling Reasoning can Improve Factuality in Large Language Models}, 
      author={Mike Zhang and Johannes Bjerva and Russa Biswas},
      year={2025},
      eprint={2505.11140},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2505.11140}, 
}

Code: https://github.com/jjzha/fs1