Pretraining large language models is a complex endeavor influenced by multiple factors, including model architecture, data quality, training continuity, and hardware constraints. In this paper, we share insights gained from the experience of training DMaS-LLaMa-Lite, a fully open source, 1.7-billion-parameter, LLaMa-based model, on approximately 20 billion tokens of carefully curated data. We chronicle the full training trajectory, documenting how evolving validation loss levels and downstream benchmarks reflect transitions from incoherent text to fluent, contextually grounded output. Beyond pretraining, we extend our analysis to include a post-training phase focused on instruction tuning, where the model was refined to produce more contextually appropriate, user-aligned responses. We highlight practical considerations such as the importance of restoring optimizer states when resuming from checkpoints, and the impact of hardware changes on training stability and throughput. While qualitative evaluation provides an intuitive understanding of model improvements, our analysis extends to various performance benchmarks, demonstrating how high-quality data and thoughtful scaling enable competitive results with significantly fewer training tokens. By detailing these experiences and offering training logs, checkpoints, and sample outputs, we aim to guide future researchers and practitioners in refining their pretraining strategies. The training script is available on Github at https://github.com/McGill-DMaS/DMaS-LLaMa-Lite-Training-Code. The model checkpoints are available on Huggingface at https://huggingface.co/collections/McGill-DMaS/dmas-llama-lite-6761d97ba903f82341954ceb.

Recent advances in large language models (LLMs) have predominantly focused on maximizing accuracy and reasoning capabilities, often overlooking crucial computational efficiency considerations. While this approach has yielded impressive accuracy improvements, it has led to methods that may be impractical for real-world deployment due to computational overhead and latency constraints. This paper investigates the potential synergy between reasoning enhancement and computational efficiency by analyzing the integration of two contrasting approaches: Quiet-STaR (Self-Taught Reasoner) and REBASE (REward BAlanced SEarch). Through comprehensive empirical analysis using the Mistral-7B model on the GSM8K dataset, we demonstrate that while each method excels in its primary objective-Quiet-STaR achieving superior accuracy (32.03%) despite high computational cost (554.66s runtime, 12.73T FLOPs), and REBASE providing exceptional efficiency (8.47s runtime, 2.35T FLOPs) while maintaining baseline-comparable accuracy (10.94%)-their integration reveals fundamental challenges in reconciling reasoning depth with computational efficiency. The combined approach unexpectedly results in degraded performance (9.38% accuracy, 143.66s runtime), highlighting critical insights about the complex interplay between reasoning enhancement and efficiency optimization in LLMs. Our findings illuminate the need for novel architectures and algorithms specifically designed to bridge the gap between these competing objectives, while providing concrete directions for future research in compute-efficient reasoning methods.

We show that current open-source foundational LLMs possess instruction capability and German legal background knowledge that is sufficient for some legal analysis in an educational context. However, model capability breaks down in very specific tasks, such as the classification of "Gutachtenstil" appraisal style components, or with complex contexts, such as complete legal opinions. Even with extended context and effective prompting strategies, they cannot match the Bag-of-Words baseline. To combat this, we introduce a Retrieval Augmented Generation based prompt example selection method that substantially improves predictions in high data availability scenarios. We further evaluate the performance of pre-trained LLMs on two standard tasks for argument mining and automated essay scoring and find it to be more adequate. Throughout, pre-trained LLMs improve upon the baseline in scenarios with little or no labeled data with Chain-of-Thought prompting further helping in the zero-shot case.

Existing large video-language models (LVLMs) struggle to comprehend long videos correctly due to limited context. To address this problem, fine-tuning long-context LVLMs and employing GPT-based agents have emerged as promising solutions. However, fine-tuning LVLMs would require extensive high-quality data and substantial GPU resources, while GPT-based agents would rely on proprietary models (e.g., GPT-4o). In this paper, we propose Video Retrieval-Augmented Generation (Video-RAG), a training-free and cost-effective pipeline that employs visually-aligned auxiliary texts to help facilitate cross-modality alignment while providing additional information beyond the visual content. Specifically, we leverage open-source external tools to extract visually-aligned information from pure video data (e.g., audio, optical character, and object detection), and incorporate the extracted information into an existing LVLM as auxiliary texts, alongside video frames and queries, in a plug-and-play manner. Our Video-RAG offers several key advantages: (i) lightweight with low computing overhead due to single-turn retrieval; (ii) easy implementation and compatibility with any LVLM; and (iii) significant, consistent performance gains across long video understanding benchmarks, including Video-MME, MLVU, and LongVideoBench. Notably, our model demonstrates superior performance over proprietary models like Gemini-1.5-Pro and GPT-4o when utilized with a 72B model.

Background: Despite the current ubiquity of Large Language Models (LLMs) across the medical domain, there is a surprising lack of studies which address their reasoning behaviour. We emphasise the importance of understanding reasoning behaviour as opposed to high-level prediction accuracies, since it is equivalent to explainable AI (XAI) in this context. In particular, achieving XAI in medical LLMs used in the clinical domain will have a significant impact across the healthcare sector. Results: Therefore, we define the concept of reasoning behaviour in the specific context of medical LLMs. We then categorise and discuss the current state of the art of methods which evaluate reasoning behaviour in medical LLMs. Finally, we propose theoretical frameworks which can empower medical professionals or machine learning engineers to gain insight into the low-level reasoning operations of these previously obscure models. Conclusion: The subsequent increased transparency and trust in medical machine learning models by clinicians as well as patients will accelerate the integration, application as well as further development of medical AI for the healthcare system as a whole

Funding and support: The Generative AI Challenge is funded by grants from the Future Health Research and Innovation Fund (FHRIF), Grant ID IC2023-GAIA/11. Conflict of interest statement: The authors declare no conflicts of interest. Abstract This exploratory pilot study investigated the potential of combining a domain-specific model, BERN2, with large language models (LLMs) to enhance automated disease phenotyping from research survey data. Motivated by the need for efficient and accurate methods to harmonize the growing volume of survey data with standardized disease ontologies, we employed BERN2, a biomedical named entity recognition and normalization model, to extract disease information from the ORIGINS birth cohort survey data. After rigorously evaluating BERN2's performance against a manually curated ground truth dataset, we integrated various LLMs using prompt engineering, Retrieval-Augmented Generation (RAG), and Instructional Fine-Tuning (IFT) to refine the model's outputs. BERN2 demonstrated high performance in extracting and normalizing disease mentions, and the integration of LLMs, particularly with Few Shot Inference and RAG orchestration, further improved accuracy. This approach, especially when incorporating structured examples, logical reasoning prompts, and detailed context, offers a promising avenue for developing tools to enable efficient cohort profiling and data harmonization across large, heterogeneous research datasets. Introduction The increasing availability of research survey data from cohort studies and clinical trials offers unprecedented opportunities to advance biomedical research and improve healthcare (1).

Aligning large language models (LLMs) to human preferences is a crucial step in building helpful and safe AI tools, which usually involve training on supervised datasets. Popular algorithms such as Direct Preference Optimization rely on pairs of AI-generated responses ranked according to human feedback. The response pair annotation process is the most labor-intensive and costly part of the alignment pipeline, and improving its efficiency and annotation quality would have a meaningful impact on AI development. We propose REAL: Response Embedding-based Alignment for LLMs, a strategy for constructing a high-quality training dataset that focuses on acquiring the most informative response pairs for labeling out of a set of response candidates. Our selection process is based on embedding responses independently of prompts. Experimental results on real-world dataset SHP2 and synthetic HH-RLHF benchmarks indicate that choosing dissimilar response pairs enhances the direct alignment of LLMs while reducing inherited labeling errors. The model aligned on dissimilar response pairs obtained a better margin and win rate on the dialogue task. Our findings suggest that focusing on distinct pairs can reduce the label error to improve the efficiency of LLM alignment, saving up to 65% of annotators' work.

Aligning large language models (LLMs) through fine-tuning is essential for tailoring them to specific applications. Therefore, understanding what LLMs learn during the alignment process is crucial. Recent studies suggest that alignment primarily adjusts a model's presentation style rather than its foundational knowledge, indicating that only certain components of the model are significantly impacted. To delve deeper into LLM alignment, we propose to identify which layers within LLMs are most critical to the alignment process, thereby uncovering how alignment influences model behavior at a granular level. We propose a novel approach to identify the important layers for LLM alignment (ILA). It involves learning a binary mask for each incremental weight matrix in the LoRA algorithm, indicating the significance of each layer. ILA consistently identifies important layers across various alignment datasets, with nearly 90% overlap even with substantial dataset differences, highlighting fundamental patterns in LLM alignment. Experimental results indicate that freezing non-essential layers improves overall model performance, while selectively tuning the most critical layers significantly enhances fine-tuning efficiency with minimal performance loss.

Efficient image tokenization with high compression ratios remains a critical challenge for training generative models. We present SoftVQ-VAE, a continuous image tokenizer that leverages soft categorical posteriors to aggregate multiple codewords into each latent token, substantially increasing the representation capacity of the latent space. When applied to Transformer-based architectures, our approach compresses 256x256 and 512x512 images using as few as 32 or 64 1-dimensional tokens. Not only does SoftVQ-VAE show consistent and high-quality reconstruction, more importantly, it also achieves state-of-the-art and significantly faster image generation results across different denoising-based generative models. Remarkably, SoftVQ-VAE improves inference throughput by up to 18x for generating 256x256 images and 55x for 512x512 images while achieving competitive FID scores of 1.78 and 2.21 for SiT-XL. It also improves the training efficiency of the generative models by reducing the number of training iterations by 2.3x while maintaining comparable performance. With its fully-differentiable design and semantic-rich latent space, our experiment demonstrates that SoftVQ-VAE achieves efficient tokenization without compromising generation quality, paving the way for more efficient generative models. Code and model are released.

This technical report presents Yi-Lightning, our latest flagship large language model (LLM). It achieves exceptional performance, ranking 6th overall on Chatbot Arena, with particularly strong results (2nd to 4th place) in specialized categories including Chinese, Math, Coding, and Hard Prompts. Yi-Lightning leverages an enhanced Mixture-of-Experts (MoE) architecture, featuring advanced expert segmentation and routing mechanisms coupled with optimized KV-caching techniques. Our development process encompasses comprehensive pre-training, supervised fine-tuning (SFT), and reinforcement learning from human feedback (RLHF), where we devise deliberate strategies for multi-stage training, synthetic data construction, and reward modeling. Furthermore, we implement RAISE (Responsible AI Safety Engine), a four-component framework to address safety issues across pre-training, post-training, and serving phases. Empowered by our scalable super-computing infrastructure, all these innovations substantially reduce training, deployment and inference costs while maintaining high-performance standards. With further evaluations on public academic benchmarks, Yi-Lightning demonstrates competitive performance against top-tier LLMs, while we observe a notable disparity between traditional, static benchmark results and real-world, dynamic human preferences. This observation prompts a critical reassessment of conventional benchmarks' utility in guiding the development of more intelligent and powerful AI systems for practical applications. Yi-Lightning is now available through our developer platform at https://platform.lingyiwanwu.com.