The temperature parameter plays a profound role during training and/or inference with large foundation models (LFMs) such as large language models (LLMs) and CLIP models. Particularly, it adjusts the logits in the softmax function in LLMs, which is crucial for next token generation, and it scales the similarities in the contrastive loss for training CLIP models. A significant question remains: "Is it viable to learn a neural network to predict a personalized temperature of any input data for enhancing LFMs?" In this paper, we present a principled framework for learning a small yet generalizable temperature prediction network (TempNet) to improve LFMs. Our solution is composed of a novel learning framework with robust losses underpinned by constrained distributionally robust optimization (DRO), and a properly designed TempNet with theoretical inspiration. TempNet can be trained together with a large foundation model from scratch or learned separately given a pretrained foundation model. It is not only useful for predicting personalized temperature to promote the training of LFMs but also generalizable and transferable to new tasks. Our experiments on LLMs and CLIP models demonstrate that TempNet greatly improves the performance of existing solutions or models.

We introduce a principled framework for developing a small yet generalizable network for temperature prediction, TempNet, aimed at enhancing large foundation models (LFMs) such as large language models (LLMs) and CLIP models. The Temperature Network is a plug-and-play architecture that can be implemented atop LFMs. Our solution is composed of a novel learning framework with robust losses underpinned by constrained distributionally robust optimization (DRO), and a properly designed TempNet with theoretical inspiration. TempNet can be trained together with a large foundation model from scratch or learned separately given a pretrained foundation model. It is not only useful for predicting personalized temperature to promote the training of LFMs but also generalizable and transferable to new tasks.

In the figure above, we present the framework of training LFMs with TempNet on the left and the structure of TempNet on the right.

Results of training LLMs in various settings, including training from scratch, finetuning a pretrained LLM model, and learning TempNet only with a frozen LLM model.

Results on contrastive learning. For image-text retrieval on Flickr30K and MSCOCO, we compute IR@1 and TR@1 for the Recall@1 on image-retrieval (IR) and text-retrieval (TR). For classification tasks, we compute top-1 accuracy (%). We report the average of scores and standard deviation over 3 runs with different random seeds.

In the following experiments, we investigate two components of our framework: the DRO-based robust loss and the role of TempNet. One can observe that both components significantly impact performance.

To test TempNet's performance in instruction-following tasks, we fix the LLaMA2 Chat models and trained TempNet, then test them on the AlpacaEval benchmark. We present the results on three different model sizes in the table below, including the training times of TempNet on Nvidia A100-80G GPUs and their win rates on AlpacaEval data. The results demonstrate that our TempNet can converge quickly and achieve consistent improvements.

Here, we reveal why TempNet enhances performance by comparing the performances of LLaMA2 7B Chat (with the default $\tau=0.7$) and LLaMA2 7B Chat + TempNet on the AlpacaEval dataset. We select a representative example for which the AlpacaEval annotator, GPT-4, deems the response from LLaMA2 + TempNet to be not only superior to that of LLaMA but also better than the baseline response generated by GPT-4.

It is a relatively subjective task to naming a dish. When the temperature value is lower, it can be observed that the LLaMA2 7B Chat model's output is relatively fixed and lacks creativity. With a higher temperature, the model generates more creative names. With TempNet, in the process of generating names for this task, LLaMA2 7B Chat produces a higher averaged temperature value of 0.82, ultimately creating a novel name Tunanadoes.

We further demonstrate the predicted temperature parameter produced by the TempNet each time a token here. One can clearly observe that when the potential possibilities for the token to be predicted are numerous, the temperature values are higher. Conversely, when there are fewer potential possibilities for the token to be predicted, the temperature values are lower.

For more details, please refer to our paper

We conduct experiments across various tasks and models to validate the effectiveness of TempNet. Given the different training frameworks required by each model, we distribute the training code for different models across four directories: GPT_2, LLaMA-1, LLaMA-2, and Bimodal-CL.

We upload the base models for LLaMA 2 Chat 7B, 13B, 70B, and their respective TempNets to Hugging Face. The tempnet.py in the repository contains the definition of the TempNet class and a class that inherits from Hugging Face's LLaMA, including TempNet. People can download this file and use the following code to perform inference with LLaMA that incorporates TempNet.

This repository benefits from ALBEF, GPT-NeoX, LLaMA, Megatron-LM, lm-evaluation-harness, Stanford Alpaca, and DeepSpeed.

Thanks for their wonderful works and their efforts to further research.

If you find this tutorial helpful, please cite our paper:

@article{qiu2024to,
  title={To Cool or not to Cool? Temperature Network Meets Large Foundation Models via DRO},
  author={Zi-Hao Qiu, Siqi Guo, Mao Xu, Tuo Zhao, Lijun Zhang, and Tianbao Yang},
  journal={arXiv preprint arXiv:2404.04575},
  year={2024}
}