Nemotron-4 15B Technical Report

We introduce Nemotron-4 15B, a 15-billion-parameter large multilingual language model trained on 8 trillion text tokens. Nemotron-4 15B demonstrates strong performance when assessed on English, multilingual, and coding tasks: it outperforms all existing similarly-sized open models on 4 out of 7 downstream evaluation areas and achieves competitive performance to the leading open models in the remaining ones. Specifically, Nemotron-4 15B exhibits the best multilingual capabilities of all similarly-sized models, even outperforming models over four times larger and those explicitly specialized for multilingual tasks.

SPRING: GPT-4 Out-performs RL Algorithms by Studying Papers and Reasoning

Open-world survival games pose significant challenges for AI algorithms due to their multi-tasking, deep exploration, and goal prioritization requirements. Despite reinforcement learning (RL) being popular for solving games, its high sample complexity limits its effectiveness in complex open-world games like Crafter or Minecraft. We propose a novel approach, SPRING, to read the game's original academic paper and use the knowledge learned to reason and play the game through a large language model (LLM). Prompted with the LaTeX source as game context and a description of the agent's current observation, our SPRING framework employs a directed acyclic graph (DAG) with game-related questions as nodes and dependencies as edges. We identify the optimal action to take in the environment by traversing the DAG and calculating LLM responses for each node in topological order, with the LLM's answer to final node directly translating to environment actions. In our experiments, we study the quality of in-context "reasoning" induced by different forms of prompts under the setting of the Crafter open-world environment. Our experiments suggest that LLMs, when prompted with consistent chain-of-thought, have great potential in completing sophisticated high-level trajectories. Quantitatively, SPRING with GPT-4 outperforms all state-of-the-art RL baselines, trained for 1M steps, without any training. Finally, we show the potential of games as a test bed for LLMs.

Plan, Eliminate, and Track -- Language Models are Good Teachers for Embodied Agents

Pre-trained large language models (LLMs) capture procedural knowledge about the world. Recent work has leveraged LLM's ability to generate abstract plans to simplify challenging control tasks, either by action scoring, or action modeling (fine-tuning). However, the transformer architecture inherits several constraints that make it difficult for the LLM to directly serve as the agent: e.g. limited input lengths, fine-tuning inefficiency, bias from pre-training, and incompatibility with non-text environments. To maintain compatibility with a low-level trainable actor, we propose to instead use the knowledge in LLMs to simplify the control problem, rather than solving it. We propose the Plan, Eliminate, and Track (PET) framework. The Plan module translates a task description into a list of high-level sub-tasks. The Eliminate module masks out irrelevant objects and receptacles from the observation for the current sub-task. Finally, the Track module determines whether the agent has accomplished each sub-task. On the AlfWorld instruction following benchmark, the PET framework leads to a significant 15% improvement over SOTA for generalization to human goal specifications.

Self-Refine: Iterative Refinement with Self-Feedback

Like people, LLMs do not always generate the best text for a given generation problem on their first try (e.g., summaries, answers, explanations). Just as people then refine their text, we introduce SELF-REFINE, a framework for similarly improving initial outputs from LLMs through iterative feedback and refinement. The main idea is to generate an output using an LLM, then allow the same model to provide multi-aspect feedback for its own output; finally, the same model refines its previously generated output given its own feedback. Unlike earlier work, our iterative refinement framework does not require supervised training data or reinforcement learning, and works with a single LLM. We experiment with 7 diverse tasks, ranging from review rewriting to math reasoning, demonstrating that our approach outperforms direct generation. In all tasks, outputs generated with SELF-REFINE are preferred by humans and by automated metrics over those generated directly with GPT-3.5 and GPT-4, improving on average by absolute 20% across tasks.

Adding Instructions during Pretraining: Effective Way of Controlling Toxicity in Language Models

Pretrained large language models have become indispensable for solving various natural language processing (NLP) tasks. However, safely deploying them in real world applications is challenging because they generate toxic content. To address this challenge, we propose two novel pretraining data augmentation strategies that significantly reduce model toxicity without compromising its utility. Our two strategies are: (1) MEDA: adds raw toxicity score as meta-data to the pretraining samples, and (2) INST: adds instructions to those samples indicating their toxicity. Our results indicate that our best performing strategy (INST) substantially reduces the toxicity probability up to 61% while preserving the accuracy on five benchmark NLP tasks as well as improving AUC scores on four bias detection tasks by 1.3%. We also demonstrate the generalizability of our techniques by scaling the number of training samples and the number of model parameters.

AutoBiasTest: Controllable Sentence Generation for Automated and Open-Ended Social Bias Testing in Language Models

Social bias in Pretrained Language Models (PLMs) affects text generation and other downstream NLP tasks. Existing bias testing methods rely predominantly on manual templates or on expensive crowd-sourced data. We propose a novel AutoBiasTest method that automatically generates sentences for testing bias in PLMs, hence providing a flexible and low-cost alternative. Our approach uses another PLM for generation and controls the generation of sentences by conditioning on social group and attribute terms. We show that generated sentences are natural and similar to human-produced content in terms of word length and diversity. We illustrate that larger models used for generation produce estimates of social bias with lower variance. We find that our bias scores are well correlated with manual templates, but AutoBiasTest highlights biases not captured by these templates due to more diverse and realistic test sentences. By automating large-scale test sentence generation, we enable better estimation of underlying bias distributions

Can You Label Less by Using Out-of-Domain Data? Active & Transfer Learning with Few-shot Instructions

Labeling social-media data for custom dimensions of toxicity and social bias is challenging and labor-intensive. Existing transfer and active learning approaches meant to reduce annotation effort require fine-tuning, which suffers from over-fitting to noise and can cause domain shift with small sample sizes. In this work, we propose a novel Active Transfer Few-shot Instructions (ATF) approach which requires no fine-tuning. ATF leverages the internal linguistic knowledge of pre-trained language models (PLMs) to facilitate the transfer of information from existing pre-labeled datasets (source-domain task) with minimum labeling effort on unlabeled target data (target-domain task). Our strategy can yield positive transfer achieving a mean AUC gain of 10.5% compared to no transfer with a large 22b parameter PLM. We further show that annotation of just a few target-domain samples via active learning can be beneficial for transfer, but the impact diminishes with more annotation effort (26% drop in gain between 100 and 2000 annotated examples). Finally, we find that not all transfer scenarios yield a positive gain, which seems related to the PLMs initial performance on the target-domain task.

Evaluating Parameter Efficient Learning for Generation

Parameter efficient learning methods (PERMs) have recently gained significant attention as they provide an efficient way for pre-trained language models (PLMs) to adapt to a downstream task. However, these conclusions are mostly drawn from in-domain evaluations over the full training set. In this paper, we present comparisons between PERMs and finetuning from three new perspectives: (1) the effect of sample and model size to in-domain evaluations, (2) generalization to unseen domains and new datasets, and (3) the faithfulness of generations. Our results show that for in-domain settings (a) there is a cross point of sample size for which PERMs will perform better than finetuning when training with fewer samples, and (b) larger PLMs have larger cross points. For cross-domain and cross-dataset cases, we show that (a) Adapter (Houlsby et al., 2019) performs the best amongst all the PERMs studied here, and (b) it outperforms finetuning if the task dataset is below a certain size. We also compare the faithfulness of generations and show that PERMs can achieve better faithfulness score than finetuning, especially for small training set, by as much as 6%. Finally, we apply Adapter to MT-NLG 530b (Smith et al., 2022) and achieve new state-of-the-art results on Xsum (Narayan et al., 2018) for all ROUGE scores (ROUGE-1 49.17, ROUGE-2 27.20, ROUGE-L 40.98).

Context Generation Improves Open Domain Question Answering

Closed-book question answering (QA) requires a model to directly answer an open-domain question without access to any external knowledge. Prior work on closed-book QA either directly finetunes or prompts a pretrained language model (LM) to leverage the stored knowledge. However, they do not fully exploit the parameterized knowledge. To address this issue, we propose a two-stage, closed-book QA framework which employs a coarse-to-fine approach to extract relevant knowledge and answer a question. Our approach first generates a related context for a given question by prompting a pretrained LM. We then prompt the same LM for answer prediction using the generated context and the question. Additionally, to eliminate failure caused by context uncertainty, we marginalize over generated contexts. Experimental results on three QA benchmarks show that our method significantly outperforms previous closed-book QA methods (e.g. exact matching 68.6% vs. 55.3%), and is on par with open-book methods that exploit external knowledge sources (e.g. 68.6% vs. 68.0%). Our method is able to better exploit the stored knowledge in pretrained LMs without adding extra learnable parameters or needing finetuning, and paves the way for hybrid models that integrate pretrained LMs with external knowledge.