UMM-SFT Full Syn-Concept LoRA Endpoints

Private research release of the currently published trained Full Syn-Concept e2 LoRA endpoints. Base-model weights are not included. Each adapter must be loaded with the corresponding model-family base checkpoint and runtime.

Contents

Model Arm Adapter subfolder Source run
BLIP3o UG blip3o/ug blip3o-syn-concept-full-ug-lora-w2-e2-fullaux-2d8cadd2c4
BLIP3o U blip3o/u blip3o-syn-concept-full-u-lora-w2-e2-fullaux-d7970824dc
BLIP3o G blip3o/g blip3o-syn-concept-full-g-lora-w2-e2-fullaux-59ea7f9147
BAGEL UG bagel/ug bagel-syn-concept-full-ug-bagel-lora-dty-w4-e2-mbs2-fullaux-2309021953
BAGEL U bagel/u bagel-syn-concept-full-u-bagel-lora-lb-gzs-w4-e2-mbs2-fullaux-92297867ef
BAGEL G bagel/g bagel-syn-concept-full-g-bagel-lora-lb-gzs-w4-e2-mbs2-fullaux-e8fd585f08
Janus-Pro UG janus/ug janus-syn-concept-full-ug-lora-w4-e2-fullaux-per-stream-yuan-mbs1-lazy-ckpt5-abb6283818
Janus-Pro U janus/u janus-syn-concept-full-u-lora-w4-e2-fullaux-per-stream-lb-gzs-mbs3-lazy-4e911fc580

Every published subfolder contains adapter_model.safetensors, adapter_config.json, and metadata.json. Family-specific tokenizer assets are included when the checkpoint exporter emitted them. The root manifest.json rolls up the file sizes, SHA-256 receipts, tensor counts, source runs, and effective training steps.

Publication Boundary

  • Scope: trained Full Syn-Concept e2 LoRA endpoints only; Tiny runs are excluded.
  • Janus-Pro includes only the new per_stream_global_batch UG and U checkpoints. Historical packed Janus checkpoints were not uploaded. The fresh per-stream G arm is still training.
  • Emu3.5 UG training is complete at step_000588, but its adapter is not yet in this repository because the Hub rejected every new LFS object with 403 Private repository storage limit reached. Emu3.5 U and G are not trained to final checkpoints.
  • Only adapter payloads are published. Base weights, optimizer/RNG state, recorder artifacts, and full checkpoints are intentionally excluded.
  • Endpoint results are descriptive comparisons, not controlled causal estimates of U-to-G or G-to-U transfer.

Janus Per-Stream Results

J / M denotes joint accuracy / attribute macro accuracy. Full-sentence and prompt-matched Syn-G remain separate evaluation surfaces.

Arm Syn-G full J / M Syn-G matched J / M Syn-U Caption J / M Syn-U MCQ J / M
UG 2.98 / 38.10 9.52 / 59.67 100 / 100 100 / 100
U 0 / 12.50 0 / 11.46 100 / 100 100 / 100

Loading

Download only the chosen subfolder, then pass that local directory to the matching family PEFT loading path. The repository does not define or redistribute model-family base checkpoints.

from pathlib import Path
from huggingface_hub import snapshot_download

root = snapshot_download(
    repo_id="xzz789/umm-sft-full-syn-concept-lora",
    allow_patterns=["janus/ug/**"],
)
adapter_dir = Path(root) / "janus" / "ug"

Use metadata.json or manifest.json to verify the downloaded adapter before loading it.

Downloads last month
-
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support