vision-adapter-manifests

The 45% agentic / 45% reasoning-doc / 10% conversational SFT mix (114,024 train rows + 2,328 held-out validation rows) used to train the Vision-Adapter project. Includes the full cauldron pull from which the mix was sampled.

The image corpus is a separate HF dataset repo (keypa/vision-adapter-images).

Contents

• train_manifest.jsonl — the actual train mixture (45% agentic / 45% doc / 10% conversational). Every row: {emb: "embeddings/<sha1>.pt", user, assistant, g: "agentic|doc|conv"}.
• train_manifest_val.jsonl — held-out validation split (never optimized).
• cauldron_manifest.jsonl — raw cauldron pull (≈1.9M rows) before the sampled recipe.