An official website of the United States government
Recently, there have been significant advances and wide-scale use of generative AI in natural language generation. Models such as OpenAI’s GPT3 and Meta’s LLaMA are widely used in chatbots, to summarize documents, and to generate creative content. These advances raise concerns about abuses of these models, especially in social media settings, such as large-scale generation of disinformation, manipulation campaigns that use AI-generated content, and personalized scams. We used stylometry (the analysis of style in natural language text) to analyze the style of AI-generated text. Specifically, we applied an existing authorship verification (AV) model that can predict if two documents are written by the same author on texts generated by GPT2, GPT3, ChatGPT and LLaMA. Our AV model was trained only on human-written text and was effectively used in social media settings to analyze cases of abuse. We generated texts by providing the language models with fanfiction snippets and prompting them to complete the rest of it in the same writing style as the original snippet. We then applied the AV model across the texts generated by the language models and the human written texts to analyze the similarity of the writing styles between these texts. We found that texts generated with GPT2 had the highest similarity to the human texts. Texts generated by GPT3 and ChatGPT were very different from the human snippet, and were similar to each other. LLaMA-generated texts had some similarity to the original snippet but also has similarities with other LLaMA-generated texts and texts from other models. We then conducted a feature analysis to identify the features that drive these similarity scores. This analysis helped us answer questions like which features distinguish the language style of language models and humans, which features are different across different models, and how these linguistic features change over different language model versions. The dataset and the source code used in this analysis have been made public to allow for further analysis of new language models.
more »
« less
CAT-LM: Training Language Models on Aligned Code and Tests
Testing is an integral but often neglected part of the software development process. Classical test generation tools such as EvoSuite generate behavioral test suites by optimizing for coverage, but tend to produce tests that are hard to understand. Language models trained on code can generate code that is highly similar to that written by humans, but current models are trained to generate each file separately, as is standard practice in natural language processing, and thus fail to consider the code- under-test context when producing a test file. In this work, we propose the Aligned Code And Tests Language Model (CAT- LM), a GPT-style language model with 2.7 Billion parameters, trained on a corpus of Python and Java projects. We utilize a novel pretraining signal that explicitly considers the mapping between code and test files when available. We also drastically increase the maximum sequence length of inputs to 8,192 tokens, 4x more than typical code generation models, to ensure that the code context is available to the model when generating test code. We analyze its usefulness for realistic applications, showing that sampling with filtering (e.g., by compilability, coverage) allows it to efficiently produce tests that achieve coverage similar to ones written by developers while resembling their writing style. By utilizing the code context, CAT-LM generates more valid tests than even much larger language models trained with more data (CodeGen 16B and StarCoder) and substantially outperforms a recent test-specific model (TeCo) at test completion. Overall, our work highlights the importance of incorporating software-specific insights when training language models for code and paves the way to more powerful automated test generation.
more »
« less
- Award ID(s):
- 1910067
- PAR ID:
- 10487734
- Publisher / Repository:
- IEEE
- Date Published:
- Journal Name:
- IEEE/ACM International Conference on Automated Software Engineering (ASE)
- Format(s):
- Medium: X
- Location:
- Luxembourg City, Luxembourg
- Sponsoring Org:
- National Science Foundation
More Like this
-
-
Summary Search‐based unit test generation, if effective at fault detection, can lower the cost of testing. Such techniques rely on fitness functions to guide the search. Ultimately, such functions represent test goals that approximate—but do not ensure—fault detection. The need to rely on approximations leads to two questions—can fitness functions produce effective tests and, if so, which should be used to generate tests?To answer these questions, we have assessed the fault‐detection capabilities of unit test suites generated to satisfy eight white‐box fitness functions on 597 real faults from the Defects4J database. Our analysis has found that the strongest indicators of effectiveness are a high level of code coverage over the targeted class and high satisfaction of a criterion's obligations. Consequently, the branch coverage fitness function is the most effective. Our findings indicate that fitness functions that thoroughly explore system structure should be used as primary generation objectives—supported by secondary fitness functions that explore orthogonal, supporting scenarios. Our results also provide further evidence that future approaches to test generation should focus on attaining higher coverage of private code and better initialization and manipulation of class dependencies.more » « less
-
Over the past few years, Large Language Models of Code (Code LLMs) have started to have a significant impact on programming practice. Code LLMs are also emerging as building blocks for research in programming languages and software engineering. However, the quality of code produced by a Code LLM varies significantly by programming language. Code LLMs produce impressive results on high-resource programming languages that are well represented in their training data (e.g., Java, Python, or JavaScript), but struggle with low-resource languages that have limited training data available (e.g., OCaml, Racket, and several others). This paper presents an effective approach for boosting the performance of Code LLMs on low-resource languages using semi-synthetic data. Our approach, called MultiPL-T, generates high-quality datasets for low-resource languages, which can then be used to fine-tune any pretrained Code LLM. MultiPL-T translates training data from high-resource languages into training data for low-resource languages in the following way. 1) We use a Code LLM to synthesize unit tests for commented code from a high-resource source language, filtering out faulty tests and code with low test coverage. 2) We use a Code LLM to translate the code from the high-resource source language to a target low-resource language. This gives us a corpus of candidate training data in the target language, but many of these translations are wrong. 3) We use a lightweight compiler to compile the test cases generated in (1) from the source language to the target language, which allows us to filter our obviously wrong translations. The result is a training corpus in the target low-resource language where all items have been validated with test cases. We apply this approach to generate tens of thousands of new, validated training items for five low-resource languages: Julia, Lua, OCaml, R, and Racket, using Python as the source high-resource language. Furthermore, we use an open Code LLM (StarCoderBase) with open training data (The Stack), which allows us to decontaminate benchmarks, train models without violating licenses, and run experiments that could not otherwise be done. Using datasets generated with MultiPL-T, we present fine-tuned versions of StarCoderBase and Code Llama for Julia, Lua, OCaml, R, and Racket that outperform other fine-tunes of these base models on the natural language to code task. We also present Racket fine-tunes for two very recent models, DeepSeek Coder and StarCoder2, to show that MultiPL-T continues to outperform other fine-tuning approaches for low-resource languages. The MultiPL-T approach is easy to apply to new languages, and is significantly more efficient and effective than alternatives such as training longer.more » « less
-
Natural language generators for task-oriented dialog should be able to vary the style of the output utterance while still effectively realizing the system dialog actions and their associated semantics. While the use of neural generation for training the response generation component of conversational agents promises to simplify the process of producing high quality responses in new domains, to our knowledge, there has been very little investigation of neural generators for task-oriented dialog that can vary their response style and we know of no experiments on models that can generate responses that are different in style from those seen during training, while still maintaining semantic fidelity to the input meaning representation. Here, we show that a model that is trained to achieve a single stylistic personality target can produce outputs that combine stylistic targets. We carefully evaluate the multivoice outputs for both semantic fidelity and for similarities to and differences from the linguistic features that characterize the original training style. We show that contrary to our predictions, the learned models do not always simply interpolate model parameters, but rather produce styles that are distinct and novel from the personalities they were trained on.more » « less
-
Modern language models have the capacity to store and use immense amounts of knowledge about real-world entities, but it remains unclear how to update such knowledge stored in model parameters. While prior methods for updating knowledge in LMs successfully inject atomic facts, updated LMs fail to make inferences based on injected facts. In this work, we demonstrate that a context distillation-based approach can both impart knowledge about entities and propagate that knowledge to enable broader inferences. Our approach consists of two stages: transfer set generation and distillation on the transfer set. We first generate a transfer set by prompting a language model to generate continuations from the entity definition. Then, we update the model parameters so that the distribution of the LM (the student) matches the distribution of the LM conditioned on the definition (the teacher) on the transfer set. Our experiments demonstrate that this approach is more effective at propagating knowledge updates than fine-tuning and other gradient-based knowledge-editing methods. Moreover, it does not compromise performance in other contexts, even when injecting the definitions of up to 150 entities at once.more » « less
- Free Publicly Accessible Full Text
-
Accepted Manuscript1.0
- Conference Paper:
- The DOI is not currently available.
