[ElmoPA]

Image-only training now uses the full joint MoT architecture with absent vector modalities and class instructions encoded through the existing frozen Qwen path. The ImageNet adapter preserves the shared graph and Lightning evaluation path, supports fully noisy image generation, and fixes an immutable split and validation context.

A separate allocated-CPU preparation command verifies exact ImageNet image counts, class coverage, source bytes and decoded RGB hashes, with explicit handling of two historical non-image extraction scripts. Class-language joining does not reread the corpus.

Validation: 96 adapter CPU tests, 43 raw-preparation fixtures, and an independent integration review against the published architecture parent passed. No production ImageNet preparation, Qwen generation, GPU training or deployment is certified by these tests.

[ElmoPA]

Warning

This pull request is not mergeable via GitHub because a downstack PR is open. Once all requirements are satisfied, merge this PR as a stack on Graphite.
Learn more

• Bind QK full-v5 recovery ops to published service #326
• Add exact QK full-v5 recovery history admission #325
• Add verified ImageNet smoke and full training admission #323
• Add native ImageNet data and generation path for the joint MoT model #322 👈 (View in Graphite)
• Add optional QK normalization, RMSNorm and SwiGLU to MoT experts #321
• Add Elmo continuous objectives and verified interruption recovery #320
• Add paired modality ablations for joint image forecast quality #265 : 1 other dependent PR (#276 )
• Train joint image, proprio and action MoT with smaller vector experts #264
• Add modular pixel video training and resilient log mirroring #262
• Optimize MoT attention and enable Flash4 BF16 defaults #261
• Add continuous-time MoT world model and offline data pipeline #260
• main

This stack of pull requests is managed by Graphite. Learn more about stacking.