Shulou(Shulou.com)11/24 Report--

Following the release of the first 100 billion parameter model OPT-175B in May this year, Meta has launched an "upgraded" OPT-IML. What's the improvement this time?

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

On December 22nd, an updated version of the model, OPT-IML (Open Pre-trained Transformer), was officially launched. Meta said it had "fine-tuned 2000 language tasks, including 175 billion parameters" and would be available free of charge for non-commercial research purposes.

Let's take a look at the performance of the updated OPT-IML.

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

Compared with the old OPT model, the average performance of OPT-IML is better than that of OPT in 14 standard NLP evaluation tasks.

On the zero learning task, the size of the two models is 7% better and the 32-shot task is 4% better and 0.4% better.

In this study, the researchers described how increasing model and benchmark size affects the impact of instruction adjustment decisions on the performance of downstream tasks.

To that end, they developed OPT-IML Bench, a sizeable instruction meta-learning (IML) benchmark that contains 2000 NLP tasks, which are classified into task categories according to the existing eight benchmarks.

In order to train OPT-IML 30B and 175B, the researchers first put forward some opinions on the instruction tuning decision applied to OPT-30B from the perspective of this framework.

OPT-IML demonstrated all three generalization skills on two scales on four evaluation benchmarks (PromptSource, FLAN, Super-NaturalInstructions, and UnifiedSKG) with different goals and input formats.

It not only significantly outperforms OPT in all benchmarks, but also outperforms existing models optimized for that particular benchmark in a very competitive manner.

In addition, OPT-IML is open source, and the editor of the Github link is also below.

Github link: https://github.com/ facebookresearch / metaseq / tree / main / projects / OPT-IML next through the paper to learn about OPT-IML.

Https://github.com/ facebookresearch / metaseq / blob / main / projects / OPT-IML / optimal_paper_v1.pdf Research methods instruction fine-tuning of large language models has become an effective way to enhance its generalization ability of zero and small samples. In this study, Meta researchers made three important additions to instruction fine-tuning.

First, they compiled a large-scale instruction fine-tuning benchmark that contained 2000 NLP tasks from eight dataset sets, classified by task type.

The researchers selectively built an evaluation split on this benchmark to test the generalization ability of three different types of models:

Includes tasks from the full retention category (tasks from fully held-out categories), retention tasks from seen types (held-out tasks from seen types), and reserved instances from seen tasks (held-out instances from seen tasks).

Instruction fine-tuning to fine-tune models to make them consistent with compliance instructions is one of the current research directions of machine learning.

There are two ways to fine-tune instructions. One focuses on fine-tuning the models of various tasks using instructions and feedback annotated manually; the other focuses on adding instructions to publicly accessible benchmarks and datasets through annotations or automatically.

In this study, Meta AI members focused on the second technology and compiled a number of publicly accessible datasets, including ways to improve OPT.

In the course of the study, Meta members proposed a similar scaling method using 1836 tasks from four benchmarks. Finally, while adjusting the entire test to break the performance limits of challenging external benchmarks such as MMLU and Big-Bench Hard (BBH), the researchers described the weights of various instruction tuning strategies that could affect downstream performance.

Multi-task learning is an expression of instruction-based fine-tuning (MTL).

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

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

Instruction-based fine-tuning, by contrast, helps us improve generalization performance for problems we've never seen before. It is achieved by combining all tasks into a concept by instructions and training them together by assigning the weights of the model on all tasks.

What is OPT? Large language models, that is, natural language processing systems with more than 100 billion parameters, have changed NLP and AI research in the past few years.

Trained in a large number of different texts, these models show amazing new abilities to generate creative texts, solve basic math problems, answer reading comprehension questions, and so on.

Although in some cases, the public can interact with these models through a paid API, complete research access is still limited to a small number of resource-rich laboratories.

This restricted access limits researchers' ability to understand how and why these large language models work, and hinders the progress of known problems such as improving their robustness and reducing biases.

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 trained on public data sets. Meta AI wants more communities to participate in understanding the basic technologies of large models.

To put it simply, Meta opens access to large-scale language models for artificial intelligence research to the public, thus democratizing artificial intelligence for large-scale model research.

Compared with the old version, according to the currently released IML version of Meta, it has been fine-tuned to perform better on natural language tasks than the old version of OPT.

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

To fine-tune, the researchers used about 2000 natural language tasks. These 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 is about 6-7 per cent more accurate than OPT in zero-time learning. In 32 times of learning, the accuracy of the model with 30 billion parameters was significantly improved, and the model with 175 billion parameters was slightly improved.

By comparison, the Meta team found that OPT-IML outperformed OPT in all benchmarks and was more competitive than other instruction-based fine-tuning models in terms of zero-sample and small-sample learning accuracy.

Reference:

Https://the-decoder.com/opt-iml-meta-releases-open-source-language-model-optimized-for-tasks/

Https://wandb.ai/telidavies/ml-news/reports/OPT-IML-Meta-Releases-New-Instruction-Tuned-OPT-Models-NLP-Task-Benchmark--VmlldzozMjAzMzc1

This article comes from the official account of Wechat: Xin Zhiyuan (ID:AI_era), edited by Joey and Xinpeng

Welcome to subscribe "Shulou Technology Information " to get latest news, interesting things and hot topics in the IT industry, and controls the hottest and latest Internet news, technology news and IT industry trends.