Skip to navigation
Model CoverageMultimodalLiquidAId1-3B

d1-3B

View as Markdown

NeMo AutoModel provides experimental typed decision readouts for LiquidAI D1 models. The example keeps the backbone frozen and trains additive answer-token rows for D1-3B or the marker head for D1 Omni. Its adaptation workflow uses text-only records; the released vision and audio modules remain frozen.

Run a Baseline

From the repository root, generate fictional training, validation, and test records:

uv run python tutorials/decision_models/decision_tutorial/data.py --output-dir data/soc_triage --seed 7

Then run the D1-3B baseline on a CUDA GPU:

uv run python examples/decision/d1_posttrain.py \
  --model d1-3b --mode base --data-dir data/soc_triage \
  --cache-dir runs/d1-cache --output-dir runs/d1-3b/base

Use --model d1-omni for Omni. The example loads D1-3B in BF16 and Omni in FP32, and pins model and tokenizer revisions. Subsequent --mode sft and --mode rl stages adapt the readout; RL requires --input-dir pointing to the selected SFT stage. These artifacts initialize readout weights and do not resume optimizer or scheduler state.

Model Reference

Model Architecture

PropertyValue
ArchitectureLfm2VlForConditionalGeneration (D1-3B); D1OmniModel (Omni)
D1-3B checkpointLiquidAI/d1-3B
D1 Omni checkpointLiquidAI/d1-omni-600M
Training scopeSingle-device, frozen-backbone text decision readout adaptation
ParallelismThese native model classes declare TP, CP, PP, and EP unsupported

Ordinary LFM2-VL checkpoints retain Hugging Face’s implementation. The publisher’s explicit D1 configuration mapping selects the D1-specific arithmetic. Omni construction resolves pinned publisher perception code and requires trust_remote_code=True; cache these dependencies before offline use. The D1 examples do not establish distributed adaptation or downstream task quality. Downloaded publisher code and checkpoints retain their own license terms.