Scaling laws are useful guides for derisking expensive training runs, as they predict performance of large models using cheaper, small-scale experiments. However, there remain gaps between current scaling studies and how language models are ultimately trained and evaluated. For instance, scaling is usually studied in the compute-optimal training regime (i.e., "Chinchilla optimal" regime). In contrast, models are often over-trained to reduce inference costs. Moreover, scaling laws mostly predict loss on next-token prediction, but models are usually compared on downstream task performance. To address both shortcomings, we create a testbed of 104 models with 0.011B to 6.9B parameters trained with various numbers of tokens on three data distributions. First, we fit scaling laws that extrapolate in both the amount of over-training and the number of model parameters. This enables us to predict the validation loss of a 1.4B parameter, 900B token run (i.e., 32$\times$ over-trained) and a 6.9B parameter, 138B token run (i.e., a compute-optimal run)––each from experiments that take 300$\times$ less compute. Second, we relate the perplexity of a language model to its downstream task performance by proposing a power law. We use this law to predict top-1 error averaged over downstream tasks for the two aforementioned models, using experiments that take 20$\times$ less compute.

Pre-trained LLMs that are further trained with image data perform well on vision-language tasks. While adding images during a second training phase effectively unlocks this capability, it is unclear how much of a gain or loss this two-step pipeline gives over VLMs which integrate images earlier into the training process. To investigate this, we train models spanning various datasets, scales, image-text ratios, and amount of pre-training done before introducing vision tokens.We then fine-tune these models and evaluate their downstream performance on a suite of vision-language and text-only tasks.We find that pre-training with a mixture of image and text data allows models to perform better on vision-language tasks while maintaining strong performance on text-only evaluations.On an average of 6 diverse tasks, we find that for a 1B model, introducing visual tokens 80\% of the way through pre-training results in a 2\% average improvement over introducing visual tokens to a fully pre-trained model.

We introduce DataComp for Language Models (DCLM), a testbed for controlled dataset experiments with the goal of improving language models. As part of DCLM, we provide a standardized corpus of 240T tokens extracted from Common Crawl, effective pretraining recipes based on the OpenLM framework, and a broad suite of 53 downstream evaluations. Participants in the DCLM benchmark can experiment with data curation strategies such as deduplication, filtering, and data mixing at model scales ranging from 412M to 7B parameters. As a baseline for DCLM, we conduct extensive experiments and find that model-based filtering is key to assembling a high-quality training set. The resulting dataset, DCLM-Baseline enables training a 7B parameter language model from scratch to 64 previous state-of-the-art in open-data language models, DCLM-Baseline represents a 6.6 percentage point improvement on MMLU while being trained with 40 Llama 3 8B on MMLU (63 natural language understanding tasks while being trained with 6.6x less compute than Llama 3 8B. Our results highlight the importance of dataset design for training language models and offer a starting point for further research on data curation.

We introduce OpenFlamingo, a family of autoregressive vision-language models ranging from 3B to 9B parameters. OpenFlamingo is an ongoing effort to produce an open-source replication of DeepMind's Flamingo models. On seven vision-language datasets, OpenFlamingo models average between 80 - 89% of corresponding Flamingo performance. This technical report describes our models, training data, hyperparameters, and evaluation suite. We share our models and code at https://github.com/mlfoundations/open_flamingo.

In-context vision and language models like Flamingo support arbitrarily interleaved sequences of images and text as input. This format not only enables few-shot learning via interleaving independent supervised (image, text) examples, but also, more complex prompts involving interaction between images, e.g.,"What do image A and image B have in common?"To support this interface, pretraining occurs over web corpora that similarly contain interleaved images+text. To date, however, large-scale data of this form have not been publicly available. We release Multimodal C4, an augmentation of the popular text-only C4 corpus with images interleaved. We use a linear assignment algorithm to place images into longer bodies of text using CLIP features, a process that we show outperforms alternatives. Multimodal C4 spans everyday topics like cooking, travel, technology, etc. A manual inspection of a random sample of documents shows that a vast majority (88%) of images are topically relevant, and that linear assignment frequently selects individual sentences specifically well-aligned with each image (80%). After filtering NSFW images, ads, etc., the resulting corpus consists of 101.2M documents with 571M images interleaved in 43B English tokens.

Massive web datasets play a key role in the success of large vision-language models like CLIP and Flamingo. However, the raw web data is noisy, and existing filtering methods to reduce noise often come at the expense of data diversity. Our work focuses on caption quality as one major source of noise, and studies how generated captions can increase the utility of web-scraped datapoints with nondescript text. Through exploring different mixing strategies for raw and generated captions, we outperform the best filtering method proposed by the DataComp benchmark by 2% on ImageNet and 4% on average across 38 tasks, given a candidate pool of 128M image-text pairs. Our best approach is also 2x better at Flickr and MS-COCO retrieval. We then analyze what makes synthetic captions an effective source of text supervision. In experimenting with different image captioning models, we also demonstrate that the performance of a model on standard image captioning benchmarks (e.g., NoCaps CIDEr) is not a reliable indicator of the utility of the captions it generates for multimodal training. Finally, our experiments with using generated captions at DataComp's large scale (1.28B image-text pairs) offer insights into the limitations of synthetic text, as well as the importance of image curation with increasing training data quantity. The synthetic captions used in our experiments are now available on HuggingFace.

Multimodal datasets are a critical component in recent breakthroughs such as Stable Diffusion and GPT-4, yet their design does not receive the same research attention as model architectures or training algorithms. To address this shortcoming in the ML ecosystem, we introduce DataComp, a testbed for dataset experiments centered around a new candidate pool of 12.8 billion image-text pairs from Common Crawl. Participants in our benchmark design new filtering techniques or curate new data sources and then evaluate their new dataset by running our standardized CLIP training code and testing the resulting model on 38 downstream test sets. Our benchmark consists of multiple compute scales spanning four orders of magnitude, which enables the study of scaling trends and makes the benchmark accessible to researchers with varying resources. Our baseline experiments show that the DataComp workflow leads to better training sets. In particular, our best baseline, DataComp-1B, enables training a CLIP ViT-L/14 from scratch to 79.2% zero-shot accuracy on ImageNet, outperforming OpenAI's CLIP ViT-L/14 by 3.7 percentage points while using the same training procedure and compute. We release DataComp and all accompanying code at www.datacomp.ai.

For robots to be generally useful, they must be able to find arbitrary objects described by people (i.e., be language-driven) even without expensive navigation training on in-domain data (i.e., perform zero-shot inference). We explore these capabilities in a unified setting: language- driven zero-shot object navigation (L-ZSON). Inspired by the recent success of open-vocabulary models for image classification, we investigate a straightforward framework, CLIP on Wheels (CoW), to adapt open-vocabulary models to this task without fine-tuning. To better evaluate L-ZSON, we introduce the Pasturebenchmark, which considers finding uncommon objects, objects described by spatial and appearance attributes, and hidden objects described relative to visible objects. We conduct an in-depth empirical study by directly deploying 22 CoW baselines across Habitat, Robothor,and Pasture. In total, we evaluate over 90k navigation episodes and find that (1) CoW baselines often struggle to leverage language descriptions but are proficient at finding uncommon objects. (2) A simple Co W, with CLIP-based object localization and classical exploration-and no additional training-matches the navigation efficiency of a state-of-the-art ZSON method trained for 500M steps on HabitatMp3d data. This same CoW provides a 15.6 percentage point improvement in success over a state-of-the-art ROBOTHOR ZSON model. 1 1 For code, data, and videos, see cow.cs.columbia.edu/

We introduce Structure from Action (SfA), a framework to discover 3D part geometry and joint parameters of unseen articulated objects via a sequence of inferred interactions. Our key insight is that 3D interaction and perception should be considered in conjunction to construct 3D articulated CAD models, especially for categories not seen during training. By selecting informative interactions, Sf A discovers parts and reveals occluded surfaces, like the inside of a closed drawer. By aggregating visual observations in 3D, Sf A accurately segments multiple parts, reconstructs part geometry, and infers all joint parameters in a canonical coordinate frame. Our experiments demonstrate that a Sf A model trained in simulation can generalize to many unseen object categories with diverse structures and to real-world objects. Empirically, Sf A outperforms a pipeline of state-of-the-art components by 25.4 3D IoU percentage points on unseen categories, while matching already performant joint estimation baselines. 1 1 For code, data, and videos, see sfa.cs.columbia.edu/

Natural language processing and 2D vision models have attained remarkable proficiency on many tasks primarily by escalating the scale of training data. However, 3D vision tasks have not seen the same progress, in part due to the challenges of acquiring high-quality 3D data. In this work, we present Objaverse-XL, a dataset of over 10 million 3D objects. Our dataset comprises deduplicated 3D objects from a diverse set of sources, including manually designed objects, photogrammetry scans of landmarks and everyday items, and professional scans of historic and antique artifacts. Representing the largest scale and diversity in the realm of 3D datasets, Objaverse-XL enables significant new possibilities for 3D vision. Our experiments demonstrate the improvements enabled with the scale provided by Objaverse-XL. We show that by training Zero123 on novel view synthesis, utilizing over 100 million multi-view rendered images, we achieve strong zero-shot generalization abilities. We hope that releasing Objaverse-XL will enable further innovations in the field of 3D vision at scale.

The conventional recipe for maximizing model accuracy is to (1) train multiple models with various hyperparameters and (2) pick the individual model which performs best on a held-out validation set, discarding the remainder. In this paper, we revisit the second step of this procedure in the context of fine-tuning large pre-trained models, where fine-tuned models often appear to lie in a single low error basin. We show that averaging the weights of multiple models fine-tuned with different hyperparameter configurations often improves accuracy and robustness. Unlike a conventional ensemble, we may average many models without incurring any additional inference or memory costs—we call the results “model soups.” When fine-tuning large pre-trained models such as CLIP, ALIGN, and a ViT-G pre-trained on JFT, our soup recipe provides significant improvements over the best model in a hyperparameter sweep on ImageNet. The resulting ViT-G model, which attains 90.94% top-1 accuracy on ImageNet, achieved a new state of the art. Furthermore, we show that the model soup approach extends to multiple image classification and natural language processing tasks, improves out-of-distribution performance, and improves zero-shot performance on new downstream tasks. Finally, we analytically relate the performance similarity of weight-averaging and logit-ensembling to flatness of the loss and confidence of the predictions, and validate this relation empirically. Code is available at https://github.com/mlfoundations/model-soups.

Open-vocabulary models like CLIP achieve high accuracy across many image classification tasks. However, there are still settings where their zero-shot performance is far from optimal. We study model patching, where the goal is to improve accuracy on specific tasks without degrading accuracy on tasks where performance is already adequate. Towards this goal, we introduce PAINT, a patching method that uses interpolations between the weights of a model before fine-tuning and the weights after fine-tuning on a task to be patched. On nine tasks where zero-shot CLIP performs poorly, PAINT increases accuracy by 15 to 60 percentage points while preserving accuracy on ImageNet within one percentage point of the zero-shot model. PAINT also allows a single model to be patched on multiple tasks and improves with model scale. Furthermore, we identify cases of broad transfer, where patching on one task increases accuracy on other tasks even when the tasks have disjoint classes. Finally, we investigate applications beyond common benchmarks such as counting or reducing the impact of typographic attacks on CLIP. Our findings demonstrate that it is possible to expand the set of tasks on which open-vocabulary models achieve high accuracy without re-training them from scratch.

We propose Continuous Scene Representations (CSR), a scene representation constructed by an embodied agent navigating within a space, where objects and their relationships are modeled by continuous valued embeddings. Our method captures feature relationships between objects, composes them into a graph structure on-the-fly, and situates an embodied agent within the representation. Our key insight is to embed pair-wise relationships between objects in a latent space. This allows for a richer representation compared to discrete relations (e.g., [SUPPORT], [NEXT-TO]) commonly used for building scene representations. CSR can track objects as the agent moves in a scene, update the representation accordingly, and detect changes in room configurations. Using CSR, we outperform state-of-the-art approaches for the challenging downstream task of visual room rearrangement, without any task specific training. Moreover, we show the learned embeddings capture salient spatial details of the scene and show applicability to real world data. A summery video and code is available at prior.allenai.org/projects/csr.

Households across the world contain arbitrary objects: from mate gourds and coffee mugs to sitars and guitars. Considering this diversity, robot perception must handle a large variety of semantic objects without additional fine-tuning to be broadly applicable in homes. Recently, zero-shot models have demonstrated impressive performance in image classification of arbitrary objects (i.e., classifying images at inference with categories not explicitly seen during training). In this paper, we translate the success of zero-shot vision models (e.g., CLIP) to the popular embodied AI task of object navigation. In our setting, an agent must find an arbitrary goal object, specified via text, in unseen environments coming from different datasets. Our key insight is to modularize the task into zero-shot object localization and exploration. Employing this philosophy, we design CLIP on Wheels (CoW) baselines for the task and evaluate each zero-shot model in both Habitat and RoboTHOR simulators. We find that a straightforward CoW, with CLIP-based object localization plus classical exploration, and no additional training , often outperforms learnable approaches in terms of success, efficiency, and robustness to dataset distribution shift. This CoW achieves 6.3% SPL in Habitat and 10.0% SPL in RoboTHOR, when tested zero-shot on all categories. On a subset of four RoboTHOR categories considered in prior work, the same CoW shows a 16.1 percentage point improvement in Success over the learnable state-of-the-art baseline.

People often use physical intuition when manipulating articulated objects, irrespective of object semantics. Motivated by this observation, we identify an important embodied task where an agent must play with objects to recover their parts. To this end, we introduce Act the Part (AtP) to learn how to interact with articulated objects to discover and segment their pieces. By coupling action selection and motion segmentation, AtP is able to isolate structures to make perceptual part recovery possible without semantic labels. Our experiments show AtP learns efficient strategies for part discovery, can generalize to unseen categories, and is capable of conditional reasoning for the task. Although trained in simulation, we show convincing transfer to real world data with no fine-tuning. A summery video, interactive demo, and code will be available at https://atp.cs.columbia.edu.

Mixed Reality (MR) is a promising interface for robot programming because it can project an immersive 3D visualization of a robot's intended movement onto the real world. MR can also support hand gestures, which provide an intuitive way for users to construct and modify robot motions. We present a Mixed Reality Head-Mounted Display (MRHMD) interface that enables end-users to easily create and edit robot motions using waypoints. We describe a user study where 20 participants were asked to program a robot arm using 2D and MR interfaces to perform two pick-and-place tasks. In the primitive task, participants created typical pickand-place programs. In the adapted task, participants adapted their primitive programs to address a more complex pickand-place scenario, which included obstacles and conditional reasoning. Compared to the 2D interface, a higher number of users were able to complete both tasks in significantly less time, and reported experiencing lower cognitive workload, higher usability, and higher naturalness with the MR-HMD interface.

3