Hugging Face
AI Platform
Open-source ML platform, model hub, Transformers library.
Watched repos
Recent activity
-
Researchers at IBM developed a method to improve the consistency of large language models (LLMs) like GPT-4.1, which can significantly impact their reliability in mission-critical applications. By analyzing an agent's past trajectories and identifying "flat" decisions, where the model is uncertain, they created a new type of guideline that helps stabilize these decisions. This approach, called consistency guidelines, can improve the Pass^5 metric, which measures the fraction of tasks an agent succeeds on all runs, by up to 22.9 percentage points. AI summary
Read more → -
Async GRPO with LoRA across HF Jobs uses a bucket, a proxy, and no NCCL, leveraging Hugging Face Jobs and Storage Buckets to train a LoRA adapter and sync only that adapter to vLLM replicas. The proxy routes each rollout to the replica that already holds its KV prefix and broadcasts adapter loads to all replicas. AI summary
Read more → -
Gradio's Workflow1111 is a single canvas that integrates various media pipelines, including text-to-image, hi-resolution fix, image-to-image, prompt matrix, VLM interrogate, detection to inpaint masks, and image-to-video, using SOTA models and operator types like fn, model, and space nodes, allowing for free parallelism and zero-code REST/MCP endpoints. AI summary
Read more → -
IBM has released the top-performing zero-shot time series forecasting model, Granite Time Series PatchTST-FM-r2, with a commercial-friendly license under Apache 2.0 and OpenMDW 1.0. The model achieves strong zero-shot performance on the GIFT-Eval benchmark, ranking #2 overall among replicable, zero-shot models, and outperforms pretrained models despite being smaller in size. AI summary
Read more → -
Researchers have identified a problem with current safety alignment methods for large language models, where models are trained to refuse entire topics rather than specific subsets that may be incompatible with a deployment's policy. This approach can lead to over-refusal on benign prompts and a lack of precision in measuring model safety. To address this, a new method called Boundary-Aware Self-Distillation for Controlled LLM Safety Refusal has been proposed, which focuses on identifying and refining the boundaries between safe and unsafe subsets of a topic. AI summary
Read more → -
NeoMME is a single Transformer encoder that processes both text tokens and raw image patches, eliminating the need for a separate pretrained vision tower or text decoder. The model achieves competitive retrieval performance at compact model sizes, with the 260M variant outperforming all evaluated models strictly below 800M parameters. AI summary
Read more → -
A coding model was trained using Reinforcement Learning from Human Feedback (RLHF) to paint watercolours using the p5.brush library, achieving loose and imperfect, handmade-looking results. The model was trained with three different reward mixes, each differing in the weight split between a human preference model (HPSv3) and a pairwise judge model, and demonstrated improvements in diversity and quality of output with the pairwise judge model. The model learned to avoid producing bad paintings and to focus on producing more reliable, high-quality paintings that are closer to the pool of reference images. AI summary
Read more → -
You can add a durable memory layer to your coding agents using funes, which indexes and retrieves session logs, allowing agents to recall previous decisions and reasoning. This is achieved locally on your machine or can be bound to a Hugging Face dataset, enabling recall across multiple machines or agents. AI summary
Read more → -
Fine-tuning a 350M model for better structured outputs in 100 GRPO steps involves training a smaller model (LFM2.5-350M) on a task-specific dataset (nvidia/Nemotron-RL-instruction_following-structured_outputs) with a custom reward function that focuses on structured output compliance. The model is fine-tuned with Group Relative Policy Optimization (GRPO) using the TRL library and achieves a 7.1% improvement in overall accuracy on the IFStruct benchmark. AI summary
Read more → -
IBM's Time Series Models on Confluent enable real-time intelligence with zero configuration, built-in governance, and efficiency, allowing for seamless integration with existing systems and workflows. The models can be used for forecasting, anomaly detection, optimization, and semantic intelligence, and are accessible through Flink SQL, with native inference and no need for data scientists or separate ML infrastructure. AI summary
Read more → -
BenchMIRT is a method for auditing LLM benchmarks at the level of individual prompts, analyzing how models perform on each question or task and estimating which underlying capabilities are most closely associated with getting it right. By applying Item Response Theory (IRT), BenchMIRT can separate multiple capabilities that may contribute to performance on the same questions, revealing a more complicated picture in some evaluations. The method can also help identify which questions in an evaluation are most informative about the capability the benchmark is trying to measure. AI summary
Read more → -
The Hugging Face team has released a new library called @huggingface/kernels, which provides 200+ WebGPU kernels for local AI inference. These kernels are optimized for various machine learning architectures and workloads, and are published as complete, versioned packages on the Hugging Face Hub. The library also includes a JavaScript loader that downloads, prepares, and runs kernels directly from the Hub. AI summary
Read more → -
The Open ASR Leaderboard has added its first Global South language, Hindi, to its multilingual tab, providing a new dataset for evaluating Automatic Speech Recognition (ASR) models. The dataset, called Monsoon, consists of over 4,800 speakers and 12 attributes, including region, age, gender, and occupation, and is designed to capture regional variation and orthographic variation in Hindi. AI summary
Read more → -
Here's a summary of the article in 2-3 plain sentences for a developer/AI-ML audience: Sentence Transformers provides a library for training and fine-tuning multi-vector embedding models, which outperform general-purpose retrievers on specific domains. These models can be finetuned from scratch or built upon existing models, and their training involves selecting a starting point, defining the dataset, loss function, and training arguments, as well as using the MultiVectorEncoderTrainer class. The library offers various loss functions, training arguments, and evaluation metrics to help users optimize their multi-vector model for specific tasks and domains. AI summary
Read more → -
The Granite 4.2 LLMs are built using a dense, decoder-only transformer architecture, pre-trained from scratch on 15T tokens with a five-phase strategy, and then fine-tuned with a multi-stage reinforcement learning pipeline. The pipeline includes agentic RL for the 8B and 30B models, which learn to act with tools in real environments. AI summary
Read more → -
A new approach to model compression, called Quantization-Aware Healing (QAH), has been introduced, which can recover the capabilities of a compressed model by distilling a full-precision teacher directly into the quantized student. This method outperforms traditional healing approaches, such as quantization-aware training (QAT), and can produce a 4-bit model that surpasses its full-precision original on 7 of 9 benchmarks. AI summary
Read more → -
gr.Workflow enables developers to create AI workflows as a graph of typed nodes, where each node is a runnable step that can be connected to others through drag-and-drop interfaces, REST APIs, or code. This allows for parallel execution of multiple models or spaces, such as generating an image, removing its background, and creating a voiceover, all within a single interface. The resulting workflow can be deployed to Hugging Face Spaces or run directly from code. AI summary
Read more → -
Hugging Face's hybrid search system for Papers with Code combines keyword search with dense vector embeddings to achieve stronger retrieval performance than either branch alone. The system separates throughput work (corpus embedding) from latency-sensitive work (online search), using Hugging Face Jobs for throughput-oriented work and Inference Endpoints for low-latency embeddings. AI summary
Read more → -
Researchers have identified a phenomenon in speech recognition models where they optimize for specific benchmark tests, rather than generalizing well to real-world audio. This "benchmark optimization" or "benchmaxxing" can lead to models reproducing incorrect reference transcripts or relying on subtle acoustic cues to identify the benchmark. To address this, the authors introduce three tests to measure benchmark optimization: a consensus disagreement probe, an orthographic switching probe, and a localizing the switches probe, which together reveal that models are using surrounding acoustic context to decide whether to follow the audio or a benchmark-specific transcription policy. AI summary
Read more → -
Hugging Face has released DSpark, a speculative decoding approach that significantly speeds up inference for large language models like LFM2.5, achieving up to 3.2x faster inference on GPUs and 2.87x on-device. DSpark works by using a lightweight draft model to produce candidate tokens, then having the target model verify them in a single forward pass, reducing function-calling latency by 57%. DSpark draft model checkpoints are available on Hugging Face, and the approach can be used with SGLang or llama.cpp. AI summary
Read more →