The 'upgraded version' of OPT-IML, Meta's large model with hundreds of billions of parameters, is here, and the complete model and code are released!

In May of this year, MetaAI officially announced the release of the ultra-large model OPT-175B based on 175 billion parameters, which is also open to all communities for free.

On December 22, an updated version of the model, OPT-IML (Open Pre-trained Transformer), was officially launched. Meta said it “fine-tuned 2,000 language tasks, including 1,750 Billions of Parameters" will also be freely available for non-commercial research purposes.

How about the performance of this updated OPT-IML? Let’s take a look at the two pictures.

This time OPT-IML created two model sizes, 30B and 175B.

Compared with the old OPT model, OPT-IML outperformed OPT on average in 14 standard NLP evaluation tasks.

The two model sizes are 7%~ better on the zero-shot learning task and 4%~ and 0.4%~ better on the 32-shot task.

In this study, researchers describe how increasing model and benchmark size affects the impact of instruction tuning decisions on downstream task performance.

To this end they developed OPT-IML Bench, a sizable instructional meta-learning (IML) benchmark containing 2000 NLP tasks based on eight existing Benchmarks are divided into task categories.

In order to train OPT-IML 30B and 175B, the researchers first proposed the instruction tuning decisions applied to OPT-30B from the perspective of this framework gained insights.

On four evaluation benchmarks (PromptSource, FLAN, Super-NaturalInstructions and UnifiedSKG) with different targets and input formats, OPT-IML demonstrates all three Generalization skills.

Not only does it significantly outperform OPT across all benchmarks, it also outperforms existing models optimized for this specific benchmark in a very competitive manner.

In addition, OPT-IML has been open sourced, and the editor has also put the Github link below~

Github link: https://github.com/facebookresearch/metaseq/tree/main/projects/OPT-IML

Let’s learn about it through the paper OPT-IML.

Paper link: https://github.com/facebookresearch/metaseq/blob/main/projects/OPT-IML/optimal_paper_v1 .pdf

Research Methods

Instructional fine-tuning of large language models has become an effective method to enhance their zero-shot and few-shot generalization capabilities. In this study, Meta researchers made three important additions to instruction fine-tuning.

First, they compiled a large-scale instruction fine-tuning benchmark containing 2,000 NLP tasks from eight dataset collections, categorized by task type.

Researchers selectively constructed evaluation splits on this benchmark to test three different types of model generalization capabilities:

Includes tasks from fully held-out categories, held-out tasks from seen types, and held-out instances from seen tasks (held- out instances from seen tasks). ​

Command fine-tuning

Fine-tune the model, To make them consistent with following instructions is one of the current research directions in machine learning.

There are two methods for instruction fine-tuning. One focuses on fine-tuning models for a variety of tasks using human-annotated instructions and feedback; the other, focuses on adding instructions via annotations or automatically to publicly accessible benchmarks and datasets.

In this study, Meta AI members focused on the second technique and compiled a number of publicly accessible datasets containing methods for improving OPT.

During the research, Meta members proposed a similar scaling method using 1836 tasks from four benchmarks. Finally, while tuning the entire test to push the performance limits of challenging external benchmarks such as MMLU and Big-Bench Hard (BBH), the researchers describe the weights of various instruction tuning strategies that may impact downstream performance.

Multi-task learning

Multi-task learning is a representation of instruction-based fine-tuning (MTL).

MTL is a popular paradigm that can improve task generalization performance when combined with similar functions that share comparable parameters or representations.

In recent years, MTL has been applied to numerous NLP scenarios, mainly focusing on improving the performance of training tasks or new domains by leveraging signals from related activities.

In contrast, instruction-based fine-tuning helps us improve generalization performance to never-before-seen problems. It does this by instructing to combine all tasks into a concept and train them together by assigning the weights of the model across all tasks.

What is OPT?

Large-scale language models, natural language processing systems with over 100 billion parameters, have transformed NLP and AI research over the past few years.

These models are trained on a vast array of diverse texts, demonstrating surprising new abilities to generate creative text, solve basic math problems, answer reading comprehension questions, and more.

While in some cases the public can interact with these models via paid APIs, full research access is still limited to a handful of well-resourced labs.

This restricted access limits researchers’ ability to understand how and why these large language models work, hindering progress in improving their robustness and mitigating known issues such as bias. .

Out of its commitment to open science, Meta AI released Open Pretrained Transformer (OPT-175B) in May this year, a model with 175 billion parameters, which can be used on public data. It is trained on the set. The reason for sharing this model is that Meta AI hopes that more communities will participate in understanding the basic technology about large models.

Simply put, Meta opens access to large-scale language models used in artificial intelligence research to the public, thereby realizing the democratization of artificial intelligence in large-scale model research.

Comparison with the old version

The IML version currently released by Meta has been fine-tuned and performs better on natural language tasks than the old version of OPT.

Typical language tasks include answering questions, summarizing text, and translating.

To fine-tune, the researchers used approximately 2,000 natural language tasks. The tasks are divided into eight NLP benchmarks (OPT-IML Bench), which are also provided by the researchers.

On average, taking the 30B and 175B models as examples, OPT-IML improves the zero-shot learning accuracy by about 6-7% compared to OPT. In 32 epochs of learning, the model with 30 billion parameters showed a significant improvement in accuracy, and the model with 175 billion parameters showed a slight improvement.

After comparison, the Meta team found that the performance of OPT-IML was better than OPT on all benchmark tests, and in terms of zero-shot and few-shot learning accuracy, it was better than other Models based on instruction fine-tuning are more competitive.

The above is the detailed content of The 'upgraded version' of OPT-IML, Meta's large model with hundreds of billions of parameters, is here, and the complete model and code are released!. For more information, please follow other related articles on the PHP Chinese website!