UMM Stage-2 and Stage-3 checkpoints

Deployable checkpoints for the UMM MLLM + DiT project.

Layout

  • stage2_covt_from3k_bcd_2k/checkpoint-{7000,8000,9000}-merged/: the three merged Stage-2 CoVT checkpoints from umm_stage2_covt_from3k_bcd_2k.
  • stage3/checkpoint-18500/merged_model/: merged Stage-3 understanding model at step 18,500.
  • stage3/checkpoint-18500/projector/embeddings_connector.bin: matching step-18,500 generation projector/embedding connector sidecar.
  • stage3/checkpoint-27000/merged_model/: merged Stage-3 model at step 27,000.
  • stage3/checkpoint-27000/projector/embeddings_connector.bin: matching step-27,000 generation projector/embedding connector sidecar.

The projector files must stay paired with the checkpoint at the same step. Raw DeepSpeed optimizer shards, RNG states, validation outputs, and unrelated checkpoints are intentionally excluded. ARTIFACT_MANIFEST.json records the original server paths and byte sizes used for this upload.

Downloads last month

-

Downloads are not tracked for this model. How to track
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support