Team 9

FinScale

Faster financial risk review.

Fine-tunes NVIDIA Nemotron-3-Nano 30B on two nodes to turn long annual reports into structured risk signals linked to the source text.

1.91×

faster on two nodes

1,099 against 576 tokens per second, at 95.3% scaling efficiency.

41%

less memory per node

Expert parallel, at 1.87 times the single-node speed.

77%

severity accuracy

Up from 38% for the base model on 152 held-out passages.

1 min

to load the 59 GB model

Down from 49 minutes reading straight from shared storage.

Figures are the team's own measurements from the hackathon weekend, as reported in their submission.

The problem

Financial analysts review long annual company reports to find the important risks. Manual triage is slow and hard to standardize.

What they built

A LoRA fine-tune of NVIDIA Nemotron-3-Nano-30B-A3B that reads a risk-factor passage from an SEC Form 10-K and returns JSON with the most material risk class, its severity and a quote copied word for word from the passage.

The team compared three layouts with identical data and settings: one node, two nodes with data parallel, and two nodes with the model’s experts split across them.

What they measured

Two nodes trained 1.91 times faster than one. Splitting the experts across the nodes kept almost that speed with peak memory of 37.8 GiB per node against 63.7 GiB, and all three runs learned the same. The nodes exchanged gradients at up to 11.4 GB/s over RDMA.

On 152 passages from four companies held out of training, risk-class accuracy rose from 67.1% to 79.6% and severity accuracy from 38.2% to 77.0%. Fully valid answers rose from 67.8% to 90.1%, and the median answer took 6.8 s against 10.7 s. Labels came from a larger teacher model, and none of the passages queued for human review had been reviewed at submission.

Fixing model loading

The trainer first read the model from shared network storage in small pieces, which took 49 minutes. Copying the whole model into memory crashed the trainer, because on the GB10 that memory is also the GPU’s. Copying one shard of about 5 GB at a time and loading from that copy cut the load to about 1 minute. The team says this mattered as much as the second node.

How they used the cluster

  • Fuzzball ran the full process as repeatable jobs: download, training, evaluation and export.
  • Shared storage kept files and results across job changes.
  • Two NVIDIA GB10 nodes, one GPU each, linked over RDMA.