Boston U’s Platpus Provides Quick, Cheap, and Powerful Refinement of LLMs, Achieving Top 1 in Open LLM Leaderboard

The rapid development of Large Language Models (LLMs) has kicked start the emergence of scaling law, with the ever-increasing size of both network architectures and training datasets, LLMs demonstrate strong generalization capability to address a wide range of natural language processing tasks.

Despites the excellent performance of these large scale models, fine-tuning them to adapt to downstream tasks usually time-consuming and computational expensive.

To address this issue, in a new paper Platypus: Quick, Cheap, and Powerful Refinement of LLMs, a Boston University research team presents Platpus, a family of fine-tuned and merged LLMs that achieves the first place in HuggingFace’s Open LLM Leaderboard by performing quick, cheap and powerful refinement of conventional LLMs.

The team summarizes their main contributions as follows:

1. Open-Platypus, a small-scale dataset that consists of a curated sub-selection of public text datasets. The dataset is focused on improving LLMs’ STEM and logic knowledge, and is made up of 11 open-source datasets.
2. A description of our process of similarity exclusion in order to reduce the size of our dataset, as well as reduce data redundancy.
3. A detailed look into the ever-present phenomenon of contamination of open LLM training sets with data contained in important LLM test sets, and a description of our training data filtering process in order to avoid this pitfall.
4. A description of our selection and merging process for our specialized fine-tuned LoRA modules.

The team starts by curating a small-scale dataset, Open-Platypus. The data selection schema is inspired by the Superficial Alignment Hypothesis, which indicates that it is possible for LLMs to achieve excellent performance with minimal training data. And the researchers learn from previous study that the base models had not yet reached saturation and high-quality input data is crucial for training good models.

By finding a balanced blend of the abovementioned three points, the team curated a content filtered, instruction tuned dataset from 11 open-source datasets, which comprises mainly of human-designed questions and 10% LLM-generated questions. To further refine the dataset, they also conduct contamination check, and filter three group of questions: all duplicate, gray-area, and similar but different.

In fine-tuning and merging stages, the proposed methodology focus on two points: 1) the effectiveness of Low Rank Approximation (LoRA) training and 2) the built-in model merging capabilities of the State-of-the-art Parameter-Efficient Fine-Tuning (PEFT) library. As a result, the proposed approach is able to conserve the strong prior of pretrained LLMs while capturing specific domain knowledge.

In their empirical study, the team compared their approach against other state-of-the-art (SOTA) models. The Platypus family achieves top performance on the global Open LLM leaderboard while reducing the required amount of the fine-tuning data and overall compute compared to other SOTA models.

Overall, this work verifies the high quality of the proposed Open-Platypus dataset, which opens opportunities to further advancement of this field.

Project page: https://platypus-llm.github.io. The paper Platypus: Quick, Cheap, and Powerful Refinement of LLMs on arXiv.

---

Author: Hecate He | Editor: Chain Zhang

---

We know you don’t want to miss any news or research breakthroughs. Subscribe to our popular newsletter Synced Global AI Weekly to get weekly AI updates.