Activation Steering: One Dial, Two Costs
Activation steering edits a model's internal activations to add or remove a concept at inference time. This explainer maps its one control knob, its two costs, and when it beats prompting.
Loading…
Topic
LLMs
660 posts about LLMs. Every summary links to the original.
- 660 posts
- 49 companies
- 5 explainers
See the idea
LLMs explainers
Antibodies as a Prediction Pipeline: Where AI Models Break Down
Antibody discovery is a three-stage prediction pipeline, and AI models that excel on familiar targets degrade sharply on unfamiliar ones. This explainer maps where that gap appears and what it costs.
Agent Memory: What Outlives the Run
Agent memory is the knowledge that survives a run, not the working state inside one. This explainer separates the two, then maps the write path, the read path, and the database limits that cap recall.
The Good Stuff
Paid picks we use ourselves. Links carry a referral code — a subscription through one earns us a commission at no extra cost to you.
Newest first
Posts
Supabase Clippy: ChatGPT for Supabase Docs
Supabase introduced “Supabase Clippy,” a ChatGPT-like interface trained on its documentation, as a hidden MVP in the Supabase docs. It responds to questions by retrieving relevant documentation and streaming GPT-3's answer in realtime. The implementation parses docs into sections, creates OpenAI embeddings, stores them in Postgres with pgvector, and injects matching documents as context for generation. The project was motivated by the growth of Supabase's product suite and documentation, plus the need to support developers unfamiliar with Postgres; the current version uses only Guides and is presented as a one-week MVP intended to gather feedback. Planned directions include adding reference docs and community discussions, caching repeated queries, thumbs-up relevance feedback, and interactive guidance.
Accelerate Large Model Training using DeepSpeed
Training large models on hardware with limited GPU memory frequently causes out-of-memory errors when using standard Distributed Data Parallel. To solve this bottleneck, Hugging Face Accelerate integrates DeepSpeed ZeRO data parallelism to shard optimizer states, gradients, and model parameters across workers. For a 900-million-parameter DeBERTa model, ZeRO Stage 2 increased the maximum per-device batch size from eight to forty while achieving a 3.5-fold training speedup over DDP without degrading accuracy or F1 score. Advanced configurations allow sequence-to-sequence chatbot finetuning and full ZeRO Stage 3 CPU offloading for models like the 1.5-billion-parameter GPT-XL, enabling batch size 16 training where DDP fails. Accelerate enables these memory optimizations through simple configuration files and the accelerate launch command with minimal or zero code modifications.
Efficient Table Pre-training without Real Data: An Introduction to TAPEX
Conventional language model pre-training relies on massive natural text corpora and general-purpose objectives like masked language modeling, leaving a substantial gap when applied to structured table tasks. To bridge this gap efficiently, TAPEX (Table Pre-training via Execution) pre-trains a language model, such as BART, by learning a neural SQL executor over a synthetic corpus. The system pairs web tables with systematically sampled executable SQL queries and gathers target execution outputs via an off-the-shelf SQL engine. During downstream fine-tuning, the model consumes natural language questions and flattened tables to predict answers. TAPEX achieved new state-of-the-art results across four benchmark datasets, including WikiTableQuestions and TabFact, while achieving a pre-training speedup of nearly 50 times over TaBERT.
Machine Learning Experts - Sasha Luccioni
Research scientist Sasha Luccioni discusses her work on evaluating the environmental, societal, and ethical impacts of machine learning models. Her research includes measuring and reducing artificial intelligence emissions through initiatives such as CodeCarbon, regional grid selection, and the BigScience carbon footprint working group. BigScience aims to make large language models transparent by curating multilingual datasets, assessing manufacturing and compute costs, and analyzing participant workloads. Luccioni also emphasizes expanding model evaluation criteria beyond standard accuracy metrics to encompass efficiency, emissions, and subset-level fairness across demographics and data features. Overall, these efforts aim to democratize artificial intelligence and encourage deliberate trade-offs between model performance and computational resource demands.
We Raised $100 Million for Open & Collaborative Machine Learning 🚀
Hugging Face announced a 100 million dollar Series C funding round led by Lux Capital, alongside participants including Sequoia, Coatue, and several angel investors. The platform hosts over 100,000 pre-trained models and 10,000 datasets spanning natural language processing, computer vision, speech, and other scientific domains for more than 10,000 customer organizations. The company plans to use the capital to expand research, open-source initiatives, tooling, and efforts focused on mitigating machine learning limitations like biases, privacy risks, and energy consumption. In addition, the organization is directing the BigScience research workshop with over 1,000 participants to train a massive open-source multilingual language model. Following organizational growth from 30 to more than 120 employees in twelve months, the remote-friendly company is actively hiring across all positions and seniority levels.
BERT 101 - State Of The Art NLP Model Explained
Developed in 2018 by Google AI Language, Bidirectional Encoder Representations from Transformers addresses the historical challenge of machines lacking contextual understanding of human language. The model relies on an encoder-only Transformer architecture pre-trained on a 3.3-billion-word corpus consisting of Wikipedia and Google BooksCorpus. Training simultaneously combines masked language modeling, which hides 15% of tokenized words to enforce bidirectional context learning, with next sentence prediction across balanced sentence pairs. Pre-trained on Cloud TPUs over four days, BERT unifies solutions for more than eleven common NLP tasks and can be fine-tuned on task-specific annotated data within minutes. Unmasking experiments demonstrate that the model can also inherit distinct societal and gender biases from its underlying training corpora when predicting professions.
Getting Started with Sentiment Analysis using Python
Modern natural language processing tools allow developers to perform sentiment analysis without deep machine learning expertise. Sentiment analysis categorizes text polarity into positive, negative, or neutral labels to extract insights from large volumes of social posts, reviews, and support tickets. Using the Hugging Face Transformers library and the Hub, practitioners can quickly run pre-trained transformer pipelines or target specific multilingual and emotion-detection models with minimal Python code. For custom requirements, developers can fine-tune models like DistilBERT using the IMDB dataset via the Trainer API or train models automatically with AutoNLP. Applying these workflows to social media data enables automated sentiment distribution analysis and visualization using pandas, matplotlib, and word clouds.
Deploy GPT-J 6B for inference using Hugging Face Transformers and Amazon SageMaker
Deploying EleutherAI's 6 billion parameter GPT-J model for production inference presents latency hurdles due to large memory footprints and lengthy startup times. Loading the model via standard methods takes up to several minutes, conflicting with strict real-time response limits such as Amazon SageMaker's 60-second threshold. To overcome this limitation, the model is serialized using PyTorch's native save mechanisms, packaged into a compressed archive with supporting assets, and stored on Amazon S3. This alternative loading workflow reduces GPT-J load times down to 7.7 seconds. An Amazon SageMaker real-time endpoint is then deployed on an NVIDIA T4 GPU instance using the Hugging Face Inference Toolkit.
Training CodeParrot 🦜 from Scratch
Training code generation models from scratch requires careful data filtering, tokenization, and compute management. To build CodeParrot, a 1.5-billion-parameter Python generation model, engineers extracted 20 million Python files from BigQuery and cleaned the resulting 180 GB corpus down to 50 GB after discovering extreme duplication. A custom GPT-2 tokenizer was trained over streamed samples, and the GPT-2 large architecture was initialized with layer-scaled, full-precision attention mechanisms. The training pipeline leveraged Hugging Face Accelerate alongside gradient checkpointing and a streaming iterable dataset designed to yield fixed-length concatenated token sequences. CodeParrot was trained on roughly 25 to 30 billion tokens and evaluated on coding tasks using the pass@k metric.
Accelerating PyTorch distributed fine-tuning with Intel technologies
Transfer learning makes CPU-based clusters a viable, cost-effective option for shorter deep learning fine-tuning jobs. To accelerate PyTorch workloads, clusters can leverage Intel Ice Lake architecture featuring AVX-512 and Vector Neural Network Instructions alongside specialized software optimizations. Implementing the Intel Extension for PyTorch and the oneAPI Collective Communications Library (oneCCL) addresses communication bottlenecks across distributed nodes. In demonstrated tests fine-tuning transformer models on GLUE tasks across Amazon EC2 c6i.16xlarge instances, two-node jobs delivered roughly a 1.7x speedup, while four-node setups achieved a 3x speedup. These performance libraries and multi-node configurations maintain consistent scaling across both smaller datasets like MRPC and larger benchmarks like QQP.
Large Language Models: A New Moore's Law?
Recent advancements in generative artificial intelligence have spurred the rapid growth of large language models like Megatron-Turing NLG 530B, which require vast financial investments and substantial energy consumption. Training these massive parameter architectures demands hundreds of multi-GPU servers and generates significant carbon footprints for relatively modest benchmark gains. Rather than relying on brute-force scaling, engineering teams can adopt smaller, more frugal architectures through knowledge distillation and transfer learning techniques. Practical alternatives include fine-tuning existing pretrained models, deploying on energy-efficient cloud infrastructure, and applying optimizations like pruning, layer fusion, and quantization. These methods deliver low-latency inference and high task accuracy while significantly decreasing hardware requirements, development time, and environmental impact.
Summer at Hugging Face
Hugging Face released several major updates across its ecosystem, expanding its public Hub repository to over 16,000 models. Platform additions include Spaces for deploying Gradio and Streamlit demo applications, automatic TensorBoard instances, and evaluation metric tracking integrated with Papers With Code leaderboards. The Transformers library gained JAX/Flax support across more than 5,000 models, improved TensorFlow implementations, and introduced the transformers.onnx export module for model conversion. In research, the BigScience project completed large-scale training of a 13-billion-parameter English decoder model on Jean Zay, while the DeDLOC method enabled training the sahajBERT Bengali model without HPC infrastructure. Accepted conference papers detailed the Datasets library, prompt data-point equivalence, prompt-based fine-tuning heuristics, and block pruning techniques that reduced BERT size by 74 percent.
Deep Learning over the Internet: Training Language Models Collaboratively
Pretraining state-of-the-art Transformer language models typically requires vast computing resources and expensive hardware clusters that are inaccessible to many researchers. Standard distributed deep learning struggles over public Internet infrastructure because consumer connections suffer from limited bandwidth, dynamic peer disconnections, and restrictive firewalls. To resolve these networking bottlenecks, Distributed Deep Learning in Open Collaborations (DeDLOC) introduces an adaptive training framework that accumulates massive batch sizes across heterogeneous volunteer devices before applying optimizer updates. The framework splits gradient vectors dynamically based on participant connection speeds and allows firewalled nodes to contribute data without performing aggregation. In a practical deployment with 40 volunteers, the 18-million-parameter sahajBERT model achieved Bengali downstream benchmark performance comparable to much larger multilingual models trained on dedicated supercomputers.
Few-shot learning in practice: GPT-Neo and the 🤗 Accelerated Inference API
Limited labeled data often prevents machine learning models from achieving high performance during traditional fine-tuning. Few-shot learning addresses this hurdle in natural language processing by supplying a short task description and a few demonstration examples at inference time. EleutherAI developed GPT-Neo, an open transformer-based model trained on the Pile dataset, which can execute few-shot tasks via the Hugging Face Accelerated Inference API. Compared to larger models like GPT-3, the 2.7B parameter version of GPT-Neo requires three to four examples alongside tuned hyperparameters like temperature and stopping sequences to generate accurate completions. Because few-shot learning relies heavily on pre-trained associations, careful output monitoring and human feedback mechanisms remain necessary to prevent automated bias.
Retrieval Augmented Generation with Huggingface Transformers and Ray
Huggingface Transformers added the Retrieval Augmented Generation architecture to combine standard sequence-to-sequence models with external document retrieval for knowledge-intensive tasks. Scaling RAG fine-tuning across multiple GPUs previously created bottlenecks because the torch.distributed implementation forced the single rank 0 worker to handle all retrieval index lookups. To resolve this synchronization bottleneck, Ray was integrated into the contextual document retrieval mechanism using stateful actor abstractions. Dedicated Ray processes separate from the training workers now load the large index and process retrieval requests concurrently. This architectural separation delivers up to a 2x speedup per retrieval call and removes the strict dependency on PyTorch for training.
Leveraging Pre-trained Language Model Checkpoints for Encoder-Decoder Models
Pre-training sequence-to-sequence transformer models incurs massive computational costs, limiting development primarily to large institutions. To mitigate these expenses, encoder-decoder architectures can be warm-started using existing pre-trained checkpoints from encoder-only or decoder-only models like BERT and GPT-2. This walkthrough details the methodology and implementation of warm-starting sequence-to-sequence architectures using Hugging Face Transformers. By utilizing the EncoderDecoderModel framework alongside Seq2SeqTrainer, practitioners can construct and fine-tune models such as BERT2BERT on datasets like CNN/DailyMail. The resulting fully trained BERT2BERT model achieves a ROUGE-2 score of 18.22 on the full evaluation set, matching competitive sequence generation baselines at a fraction of standard pre-training costs.
Transformer-based Encoder-Decoder Models
Natural language generation tasks map input sequences to target sequences whose lengths cannot be known in advance and vary by content. Standard deep neural networks struggle with these variable mappings because their inputs and targets require fixed-dimensional vector representations. While recurrent neural networks addressed this challenge by generating target sequences auto-regressively from a compressed context state, transformer-based encoder-decoder architectures became the standard paradigm in modern natural language processing. The framework processes sequence-to-sequence problems by encoding source sequences and auto-regressively decoding target sequences token by token using conditional probability distributions. A step-by-step implementation demonstrates greedy decoding with Hugging Face Transformers, passing cached encoder hidden states alongside previously generated target tokens to iteratively generate German translations from English inputs.
The Reformer - Pushing the limits of language modeling
Standard transformer models hit memory bottlenecks on long sequence modeling tasks due to the quadratic asymptotic memory complexity of global self-attention and oversized positional embedding matrices. The Reformer architecture overcomes these constraints to train sequences of up to half a million tokens using under 8GB of RAM. It re-engineers transformer operations using local and Locality Sensitive Hashing self-attention, chunked feed forward layers, reversible residual layers, and axial positional encodings. In empirical benchmarks using google/reformer-crime-and-punishment, axial positional encodings reduce the model parameter count from over 136 million to approximately 2.58 million by factorizing the positional dimensions. This architectural change cuts inference memory consumption from 959 MB down to 447 MB for evaluated benchmark workloads.
How to generate text: using different decoding methods for language generation with Transformers
Auto-regressive language generation relies heavily on decoding methods to convert predicted next-token probability distributions into coherent text. Greedy search selects the single most probable token at each timestep, but it frequently leads to repetitive phrasing and misses high-probability sequences hidden behind low-probability intermediate words. Beam search alleviates this issue by maintaining multiple parallel hypotheses and can be paired with n-gram penalties to reduce immediate repetitions. Sampling approaches, including top-K and top-p sampling, offer alternative paths that can yield more fluent, varied open-ended generation. Using the Hugging Face transformers library, practitioners can configure these strategies via parameters on the model generate function.
How to train a new language model from scratch using Transformers and Tokenizers
Recent library enhancements simplify pretraining custom language models from raw text corpora. A small 84-million-parameter RoBERTa architecture comprising six layers, twelve attention heads, and a hidden dimension of 768 was pretrained on a 3 GB Esperanto text dataset. Training began by fitting a 52,000-vocabulary byte-level Byte-pair encoding tokenizer, which natively handled Esperanto diacritics and reduced average sequence lengths by approximately 30 percent compared to GPT-2 tokenization. The model was trained using masked language modeling, evaluated using fill-mask pipelines, and subsequently fine-tuned for part-of-speech tagging via standard token classification scripts. Final model artifacts and configurations were uploaded for community reuse via pre-trained model interfaces.
Related topics
All topics →Weekly digest
The best engineering reads, once a week.