SDS -- See it, Do it, Sorted: Quadruped Skill Synthesis from Single Video Demonstration

In this paper, we present SDS (``See it. Do it. Sorted.''), a novel pipeline for intuitive quadrupedal skill learning from a single demonstration video. Leveraging the Visual capabilities of GPT-4o, SDS processes input videos through our novel chain-of-thought promoting technique (SUS) and generates executable reward functions (RFs) that drive the imitation of locomotion skills, through learning a Proximal Policy Optimization (PPO)-based Reinforcement Learning (RL) policy, using environment information from the NVIDIA IsaacGym simulator. SDS autonomously evaluates the RFs by monitoring the individual reward components and supplying training footage and fitness metrics back into GPT-4o, which is then prompted to evolve the RFs to achieve higher task fitness at each iteration. We validate our method on the Unitree Go1 robot, demonstrating its ability to execute variable skills such as trotting, bounding, pacing and hopping, achieving high imitation fidelity and locomotion stability. SDS shows improvements over SOTA methods in task adaptability, reduced dependence on domain-specific knowledge, and bypassing the need for labor-intensive reward engineering and large-scale training datasets. Additional information and the open-sourced code can be found in: https://rpl-cs-ucl.github.io/SDSweb

Better Alignment with Instruction Back-and-Forth Translation

We propose a new method, instruction back-and-forth translation, to construct high-quality synthetic data grounded in world knowledge for aligning large language models (LLMs). Given documents from a web corpus, we generate and curate synthetic instructions using the backtranslation approach proposed by Li et al.(2023a), and rewrite the responses to improve their quality further based on the initial documents. Fine-tuning with the resulting (backtranslated instruction, rewritten response) pairs yields higher win rates on AlpacaEval than using other common instruction datasets such as Humpback, ShareGPT, Open Orca, Alpaca-GPT4 and Self-instruct. We also demonstrate that rewriting the responses with an LLM outperforms direct distillation, and the two generated text distributions exhibit significant distinction in embedding space. Further analysis shows that our backtranslated instructions are of higher quality than other sources of synthetic instructions, while our responses are more diverse and complex than those obtained from distillation. Overall we find that instruction back-and-forth translation combines the best of both worlds -- making use of the information diversity and quantity found on the web, while ensuring the quality of the responses which is necessary for effective alignment.

DataComp-LM: In search of the next generation of training sets for language models

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% 5-shot accuracy on MMLU with 2.6T training tokens. Compared to MAP-Neo, the 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% less compute. Our baseline model is also comparable to Mistral-7B-v0.3 and Llama 3 8B on MMLU (63% & 66%), and performs similarly on an average of 53 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.

Language models scale reliably with over-training and on downstream tasks

Scaling laws are useful guides for developing language models, but there are still 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); however, in practice, models are often over-trained to reduce inference costs. Moreover, scaling laws mostly predict loss on next-token prediction, but ultimately models are compared based on downstream task performance. In this paper, we address both shortcomings. To do so, 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 investigate scaling in the over-trained regime. We fit scaling laws that extrapolate in both the number of model parameters and the ratio of training tokens to 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$\unicode{x2014}$each from experiments that take 300$\times$ less compute. Second, we relate the perplexity of a language model to its downstream task performance via 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. Our experiments are available at https://github.com/mlfoundations/scaling.

Towards Connecting Use Cases and Methods in Interpretable Machine Learning

Despite increasing interest in the field of Interpretable Machine Learning (IML), a significant gap persists between the technical objectives targeted by researchers' methods and the high-level goals of consumers' use cases. In this work, we synthesize foundational work on IML methods and evaluation into an actionable taxonomy. This taxonomy serves as a tool to conceptualize the gap between researchers and consumers, illustrated by the lack of connections between its methods and use cases components. It also provides the foundation from which we describe a three-step workflow to better enable researchers and consumers to work together to discover what types of methods are useful for what use cases. Eventually, by building on the results generated from this workflow, a more complete version of the taxonomy will increasingly allow consumers to find relevant methods for their target use cases and researchers to identify applicable use cases for their proposed methods.

A Learning Theoretic Perspective on Local Explainability

In this paper, we explore connections between interpretable machine learning and learning theory through the lens of local approximation explanations. First, we tackle the traditional problem of performance generalization and bound the test-time accuracy of a model using a notion of how locally explainable it is. Second, we explore the novel problem of explanation generalization which is an important concern for a growing class of finite sample-based local approximation explanations. Finally, we validate our theoretical results empirically and show that they reflect what can be seen in practice.

Differentially Private Meta-Learning

Parameter-transfer is a well-known and versatile approach for meta-learning, with applications including few-shot learning, federated learning, and reinforcement learning. However, parameter-transfer algorithms often require sharing models that have been trained on the samples from specific tasks, thus leaving the task-owners susceptible to breaches of privacy. We conduct the first formal study of privacy in this setting and formalize the notion of task-global differential privacy as a practical relaxation of more commonly studied threat models. We then propose a new differentially private algorithm for gradient-based parameter transfer that not only satisfies this privacy requirement but also retains provable transfer learning guarantees in convex settings. Empirically, we apply our analysis to the problem of federated learning with personalization and show that allowing the relaxation to task-global privacy from the more commonly studied notion of local privacy leads to dramatically increased performance in recurrent neural language modeling.