Architecture: Llama decoder-only (~29M params, hidden 512, 8 layers, GQA)
Vocabulary: ~6,251 domain-specific financial transaction tokens
Trained ~3,000 steps on IBM TabFormer with NeMo AutoModel
Load with AutoModelForCausalLM.from_pretrained(...) and extract last-token embeddings

This Hub mirror exists so Spaces / Transformers can load the Git LFS checkpoint published under models/decoder-foundation-model/ in the upstream GitHub repo.

simonykq/nvidia-transaction-decoder-fm

NVIDIA Transaction Decoder Foundation Model (~29M)

simonykq/
nvidia-transaction-decoder-fm