한국어EN
Training Data & Post-TrainingKandpal et al.

Common Pile v0.1

Made by AIWritten by AI (Claude) from the source below. Check anything important against the source.How it's made

Common Pile v0.1 — 8TB of openly licensed text

Why it matters

The data-side response to litigation risk; reached Llama 1 and 2 7B-class performance with openly licensed data alone

Source — Primary sources first; secondary reports are labelled “Press”

Paperarxiv.orghttps://arxiv.org/abs/2506.05209

Added in Week 35, 2026 · See that week

On the map — ⑫ Training Data line

  1. Previous stopUS Copyright Office
  2. This recordCommon Pile v0.1
  3. Next stopPsyche shifts from proof to performance
See it on the map

Same week

Apple researchers publish The Illusion of ThinkingPaperXiang, Wu et al. publish When to use Graphs in RAG and release GraphRAG-BenchPaperAlibaba Qwen releases the Qwen3 Embedding and Reranker series (0.6B–8B)OfficialAMD makes its first MLPerf Training submissionOfficial

Nearby on the same axis

Nous Research launches Psyche Network and starts pretraining Consilience 40BOfficialUS Copyright Office releases pre-publication Copyright and AI Part 3GovernmentNous Research releases Hermes-3-Dataset (959,000 samples) on Hugging FaceOfficialA market for RL environments takes shapePress