CoolFace
Apppublic

simonykq/gl-foundation-model-demo

sourceHugging Faceupdated 1mo agoView on Hugging Face
0likes
App README

Transaction Foundation Model Demo

Status: staged as a Static Space while a ZeroGPU community grant is pending (new HF accounts need PRO or a 30-day wait for free ZeroGPU). Gradio app code (app.py, txn_tokenizer.py, requirements.txt) is in the repo — flip frontmatter to sdk: gradio + ZeroGPU once hardware is approved.

Gradio + ZeroGPU demo of the NVIDIA Transaction Decoder Foundation Model (~29M Llama decoder) from:

  • —Blog: https://developer.nvidia.com/blog/build-your-own-transaction-foundation-model-for-financial-intelligence/
  • —Upstream checkpoint (Git LFS): https://github.com/NVIDIA-AI-Blueprints/transaction-foundation-model/tree/main/models/decoder-foundation-model
  • —Hugging Face mirror: https://huggingface.co/simonykq/nvidia-transaction-decoder-fm

There is no official huggingface.co/nvidia/... model page for this checkpoint — NVIDIA ships it as a Transformers-format directory inside the GitHub repo. This Space loads the Hub mirror above via AutoModelForCausalLM.from_pretrained.

What it does

  1. 1.Domain-tokenize card transaction histories (~12 tokens / txn)
  2. 2.Extract 512-d last-token embeddings on GPU
  3. 3.Score behavioral rarity / nearest synthetic profiles

Override the model with:

bash
hf spaces variables add simonykq/gl-foundation-model-demo --env TFM_MODEL_ID=simonykq/nvidia-transaction-decoder-fm