umm-sft-ug-lora

UMM-SFT U/G LoRA archive

Release status: complete. 54 / 54 adapter records are available. The eight existing adapters remain in the now-public xzz789/umm-sft-full-syn-concept-lora; new adapters live here. This catalog links both repositories without re-uploading existing weights.

Scope and loading

Result-associated checkpoints for Synthetic Concept Full, Synthetic Concept Tiny, and KiVA v3. Different tasks, prompt surfaces, epochs and arithmetic merges are separate records. Vanilla has no LoRA. This is not an archive of every training step. Old packed Full Janus and qualification-only runs are excluded.

Download the subfolder identified in manifest.json using huggingface_hub.snapshot_download. Obtain the base and tokenizer separately from the corresponding upstream model, then use the original family-native UMM-SFT assembly and LoRA attachment scope. These adapters are not standalone AutoModel checkpoints. Janus attaches to its language model; BAGEL and BLIP3o contain multiple model regions. BLIP3o DiT readouts require the separately documented U- or G-adapted trunk. Emu3.5 results used 4-bit base inference.

Verification and interpretation

New weight files use source SHA-256, LoRA-only tensor header checks, object-store transfer and Hub LFS size/SHA-256 verification. Existing adapters use their pinned Hub revision and LFS hashes. No training, inference or new evaluation occurs for this release. Diagnostic adapters can intentionally produce poor or zero scores. Joint U/G training is different from U+G parameter addition, and parameter addition is not evidence of positive cross-task transfer. A file being available does not validate unmeasured scores.

Base weights, datasets, raw evaluation examples, optimizer/RNG state and machine-specific training configs are excluded. No blanket license replaces upstream model terms: BAGEL, BLIP3o, Janus-Pro, Emu3.5.

Published adapters