Retrieval-augmented generation (RAG) has recently become a very popular task for Large Language Models (LLMs). Evaluating them on multi-turn RAG conversations, where the system is asked to generate a response to a question in the context of a preceding conversation is an important and often overlooked task with several additional challenges. We present MTRAG: an end-to-end human-generated multi-turn RAG benchmark that reflects several real-world properties across diverse dimensions for evaluating the full RAG pipeline. MTRAG contains 110 conversations averaging 7.7 turns each across four domains for a total of 842 tasks. We also explore automation paths via synthetic data and LLM-as-a-Judge evaluation. Our human and automatic evaluations show that even state-of-the-art LLM RAG systems struggle on MTRAG. We demonstrate the need for strong retrieval and generation systems that can handle later turns, unanswerable questions, non-standalone questions, and multiple domains. MTRAG is available at https://github.com/ibm/mt-rag-benchmark.

Language models are now deployed in a wide variety of user-facing applications, often for specific purposes like answering questions about documentation or acting as coding assistants. As these models are intended for particular purposes, they should not be able to answer irrelevant queries like requests for poetry or questions about physics, or even worse, queries that can only be answered by humans like sensitive company policies. Instead we would like them to only answer queries corresponding to desired behavior and refuse all other requests, which we refer to as scoping. We find that, despite the use of system prompts, two representative language models can be poorly scoped and respond to queries they should not be addressing. We then conduct a comprehensive empirical evaluation of methods which could be used for scoping the behavior of language models. Among many other results, we show that a recently-proposed method for general alignment, Circuit Breakers (CB), can be adapted to scope language models to very specific tasks like sentiment analysis or summarization or even tasks with finer-grained scoping (e.g. When compared to standard methods like fine-tuning or preference learning, CB is more robust both for out of distribution tasks, and to adversarial prompting techniques. We also show that layering SFT and CB together often results in the best of both worlds: improved performance only on relevant queries, while rejecting irrelevant ones. In the past few years Large Language Models have exploded into the popular conscience. One major recent addition is the "alignment" process through Reinforcement Learning with Human Feedback (RLHF) (Christiano et al., 2017; Ouyang et al., 2022), which has made the current generation of language models much less likely to emit toxic content than previous generations (Wolf et al., 2017), and thus much more acceptable for general use. As a result, many businesses and individuals feel more comfortable using these technologies than they would be in the past. As a result, we have generally capable language models which refuse to answer toxic or dangerous queries, but it is still difficult to deploy these language models. Even though they may not emit toxic content as often, they still will happily answer any question, irrelevant or not. This becomes a problem when we wish to deploy language models as products in specific contexts: e.g. While language models have general language capability, there is still a need to scope them for specific uses. David Yunis is a PhD student at the Toyota Technological Institute at Chicago. Work was performed during an internship at IBM. Arrows indicate the direction of best performance.

In this work, we focus our attention on developing a benchmark for instruction-following where it is easy to verify both task performance as well as instruction-following capabilities. We adapt existing knowledge benchmarks and augment them with instructions that are a) conditional on correctly answering the knowledge task or b) use the space of candidate options in multiple-choice knowledge-answering tasks. This allows us to study model characteristics, such as their change in performance on the knowledge tasks in the presence of answer-modifying instructions and distractor instructions. In contrast to existing benchmarks for instruction following, we not only measure instruction-following capabilities but also use LLM-free methods to study task performance. We study a series of openly available large language models of varying parameter sizes (1B-405B) and closed source models namely GPT-4o-mini, GPT-4o. We find that even large-scale instruction-tuned LLMs fail to follow simple instructions in zero-shot settings. We release our dataset, the benchmark, code, and results for future work.

Growing concerns over negligent or malicious uses of AI have increased the appetite for tools that help manage the risks of the technology. In 2018, licenses with behaviorial-use clauses (commonly referred to as Responsible AI Licenses) were proposed to give developers a framework for releasing AI assets while specifying their users to mitigate negative applications. As of the end of 2023, on the order of 40,000 software and model repositories have adopted responsible AI licenses licenses. Notable models licensed with behavioral use clauses include BLOOM (language) and LLaMA2 (language), Stable Diffusion (image), and GRID (robotics). This paper explores why and how these licenses have been adopted, and why and how they have been adapted to fit particular use cases. We use a mixed-methods methodology of qualitative interviews, clustering of license clauses, and quantitative analysis of license adoption. Based on this evidence we take the position that responsible AI licenses need standardization to avoid confusing users or diluting their impact. At the same time, customization of behavioral restrictions is also appropriate in some contexts (e.g., medical domains). We advocate for ``standardized customization'' that can meet users' needs and can be supported via tooling.

The BigCode community, an open-scientific collaboration working on the responsible development of Large Language Models for Code (Code LLMs), introduces StarCoder and StarCoderBase: 15.5B parameter models with 8K context length, infilling capabilities and fast large-batch inference enabled by multi-query attention. StarCoderBase is trained on 1 trillion tokens sourced from The Stack, a large collection of permissively licensed GitHub repositories with inspection tools and an opt-out process. We fine-tuned StarCoderBase on 35B Python tokens, resulting in the creation of StarCoder. We perform the most comprehensive evaluation of Code LLMs to date and show that StarCoderBase outperforms every open Code LLM that supports multiple programming languages and matches or outperforms the OpenAI code-cushman-001 model. Furthermore, StarCoder outperforms every model that is fine-tuned on Python, can be prompted to achieve 40\% pass@1 on HumanEval, and still retains its performance on other programming languages. We take several important steps towards a safe open-access model release, including an improved PII redaction pipeline and a novel attribution tracing tool, and make the StarCoder models publicly available under a more commercially viable version of the Open Responsible AI Model license.

Prompting with natural language instructions has recently emerged as a popular method of harnessing the capabilities of large language models. Given the inherent ambiguity present in natural language, it is intuitive to consider the possible advantages of prompting with less ambiguous prompt styles, such as the use of pseudo-code. In this paper we explore if prompting via pseudo-code instructions helps improve the performance of pre-trained language models. We manually create a dataset of pseudo-code prompts for 132 different tasks spanning classification, QA and generative language tasks, sourced from the Super-NaturalInstructions dataset. Using these prompts along with their counterparts in natural language, we study their performance on two LLM families - BLOOM and CodeGen. Our experiments show that using pseudo-code instructions leads to better results, with an average increase (absolute) of 7-16 points in F1 scores for classification tasks and an improvement (relative) of 12-38% in aggregate ROUGE-L scores across all tasks. We include detailed ablation studies which indicate that code comments, docstrings, and the structural clues encoded in pseudo-code all contribute towards the improvement in performance. To the best of our knowledge our work is the first to demonstrate how pseudo-code prompts can be helpful in improving the performance of pre-trained LMs.

Large language models (LLMs) have been shown to be able to perform new tasks based on a few demonstrations or natural language instructions. While these capabilities have led to widespread adoption, most LLMs are developed by resource-rich organizations and are frequently kept from the public. As a step towards democratizing this powerful technology, we present BLOOM, a 176B-parameter open-access language model designed and built thanks to a collaboration of hundreds of researchers. BLOOM is a decoder-only Transformer language model that was trained on the ROOTS corpus, a dataset comprising hundreds of sources in 46 natural and 13 programming languages (59 in total). We find that BLOOM achieves competitive performance on a wide variety of benchmarks, with stronger results after undergoing multitask prompted finetuning. To facilitate future research and applications using LLMs, we publicly release our models and code under the Responsible AI License.

In this paper we explore the task of modeling semi-structured object sequences; in particular, we focus our attention on the problem of developing a structure-aware input representation for such sequences. Examples of such data include user activity on websites, machine logs, and many others. This type of data is often represented as a sequence of sets of key-value pairs over time and can present modeling challenges due to an ever-increasing sequence length. We propose a two-part approach, which first considers each key independently and encodes a representation of its values over time; we then self-attend over these value-aware key representations to accomplish a downstream task. This allows us to operate on longer object sequences than existing methods. We introduce a novel shared-attention-head architecture between the two modules and present an innovative training schedule that interleaves the training of both modules with shared weights for some attention heads. Our experiments on multiple prediction tasks using real-world data demonstrate that our approach outperforms a unified network with hierarchical encoding, as well as other methods including a record-centric representation and a flattened representation of the sequence.

Corresponding authors (denoted by) can be contacted at contact@bigcode-project.org The BigCode project is an open-scientific collaboration working on the responsible development of large language models for code. This tech report describes the progress of the collaboration until December 2022, outlining the current state of the Personally Identifiable Information (PII) redaction pipeline, the experiments conducted to de-risk the model architecture, and the experiments investigating better preprocessing methods for the training data. We train 1.1B parameter models on the Java, JavaScript, and Python subsets of The Stack (Kocetkov et al., 2022) and evaluate them on the MultiPL-E text-to-code benchmark (Cassano et al., 2022). We find that more aggressive filtering of near-duplicates can further boost performance and, surprisingly, that selecting files from repositories with 5+ GitHub stars deteriorates performance significantly. Our best model outperforms previous open-source multilingual code generation models (InCoder-6.7B and CodeGen-Multi-2.7B) in both left-to-right generation and infilling on the Java, JavaScript, and Python portions of MultiPL-E, despite being a substantially smaller model. All models are released under an OpenRAIL license at https://hf.co/bigcode. Over the last two years, we have witnessed tremendous progress in the development of code generating AI assistants (Chen et al., 2021; Chowdhery et al., 2022; Nijkamp et al., 2022; Fried et al., 2022; Li et al., 2022; Athiwaratkun et al., 2022). Machine learning models are now capable of assisting professional developers through the synthesis of novel code snippets, not only from surrounding code fragments, but also from natural language instructions. The models powering these code completion systems are usually referred to as Large Language Models for Code--or code LLMs--and are created by training large transformer neural networks (Vaswani et al., 2017) on big corpora of source code. However, with the exception of a few small-scale efforts (Xu et al., 2022b), there is generally a lack of transparency on the development of code LLMs, in part due to their commercial value and the legal uncertainty around distributing training data and models. Some groups have released model weights (Fried et al., 2022; Nijkamp et al., 2022) or provided access to the model through a paid API service (Chen et al., 2021; Athiwaratkun et al., 2022), but these works did not release the full training data or the preprocessing methods that were used.

Traditional systems designed for task oriented dialog utilize knowledge present only in structured knowledge sources to generate responses. However, relevant information required to generate responses may also reside in unstructured sources, such as documents. Recent state of the art models such as HyKnow and SeKnow aimed at overcoming these challenges make limiting assumptions about the knowledge sources. For instance, these systems assume that certain types of information, such as a phone number, is always present in a structured knowledge base (KB) while information about aspects such as entrance ticket prices, would always be available in documents. In this paper, we create a modified version of the MutliWOZ-based dataset prepared by SeKnow to demonstrate how current methods have significant degradation in performance when strict assumptions about the source of information are removed. Then, in line with recent work exploiting pre-trained language models, we fine-tune a BART based model using prompts for the tasks of querying knowledge sources, as well as, for response generation, without making assumptions about the information present in each knowledge source. Through a series of experiments, we demonstrate that our model is robust to perturbations to knowledge modality (source of information), and that it can fuse information from structured as well as unstructured knowledge to generate responses.