LoRA fine-tunes a small set of additional parameters while keeping most pretrained transformer weights frozen. The trained adapter weights are saved separately and later loaded on top of the same base model for task-specific inference.
AI summary
Methods worth understanding
New AI methods, the evidence behind them, and how to put them to work.
Latest weekly report
Methods worth understanding, their results, and practical limitations.
LoRA fine-tunes a small set of additional parameters while keeping most pretrained transformer weights frozen. The trained adapter weights are saved separately and later loaded on top of the same base model for task-specific inference.
AI summary
Quantize float32 sentence/document embeddings into (1) binary (1-bit-per-dimension) for very fast approximate search via Hamming distance, and optionally (2) scalar int8 embeddings (8-bit-per-dimension) for a compact candidate store. Then use a rescoring step: compare the float32 query embedding against the retrieved candidates (binary or int8) using dot-product to recover most of the original retrieval quality while keeping indices small.
AI summary
Train a single Transformer agent to imitate expert behavior by autoregressively predicting interleaved next-step embeddings of observations (and associated rewards) and actions, using a causal mask. Modality-specific encoders/decoders handle different data types (e.g., images via a CNN, continuous vectors via a linear layer, discrete values via projection), and the training loss averages per-element losses computed according to each modality (e.g., MSE for continuous/image targets and cross-entropy for discrete targets).
AI summary
Build a single inference endpoint handler that chains an ASR model, an optional diarization pipeline over ASR outputs, and an optional speculative decoding path using an assistant (smaller) model whose decoder matches the main model’s decoder architecture.
AI summary
Use a scheduler that implements rectified-flow-matching sampling for latent diffusion inference. The approach is described as using a rectified flow objective to connect data and noise distributions along a straight-line path, paired with an Euler-style discrete step formulation. The scheduler also supports a resolution-dependent timestep shift parameter to better handle noise scaling at higher resolutions.
AI summary
Fine-tune a vision-language model by framing each training example as a sequence-to-sequence generation problem where the text input is built from a task-specific prefix plus the question/prompt, paired with the corresponding image, and the target text is the desired answer/label.
AI summary
Use an interpreter that parses LLM-produced Python into an abstract syntax tree (AST) and executes nodes only if they are explicitly authorized. The system blocks or skips operations not present in allowlists (e.g., specific imports, built-in functions, and tool calls), and halts when it encounters disallowed operations or safety limits.
AI summary
Convert zero-shot VQA scoring into an LLM-driven answer-rating task: provide a question, a set of expert reference answers, and a candidate answer, then ask the LLM to output a single rating on a small discrete scale along with a rationale. The final evaluation score is derived from the selected rating.
AI summary
Quantize the weight parameters of Transformer-based diffusion components in a Diffusers pipeline using Quanto’s quantization utilities, then freeze the modules to use the quantized weights during inference. The approach supports quantizing the diffusion backbone and optionally one or more text encoder modules, with optional layer exclusion (e.g., leaving the final projection layer out for INT4).
AI summary
Use a Straight-Through Estimator to enable backpropagation through quantization steps that include non-differentiable rounding. In BitLinear-style ternary weight quantization, the forward pass rounds weights to discrete values (e.g., clamped {-1, 0, 1}); during backpropagation, the STE approximates the gradient of the rounding operation as if it were an identity function so gradients can still update the underlying full-precision weights.
AI summary
Load the SD3.5 Large transformer in 4-bit NF4 quantized form (via BitsAndBytesConfig), plug it into the StableDiffusion3Pipeline, and use model CPU offload to reduce GPU memory usage during inference.
AI summary
Use a data-centric annotation UI to import a public dataset from a Hub repository into an annotation workspace, automatically generate an initial configuration from the dataset features, define feedback questions (e.g., labels, ratings, rankings, free-form text), and then collect human feedback through the same interface for downstream fine-tuning or evaluation workflows.
AI summary