Fine-tuning Risks at a Bank | Generated by AI

Home 2026.06

Question: Is it a good idea to fine-tune GLM-5.2 or DeepSeek V4 as a base model on the bank’s finance data, then run it on public RunPod from a personal laptop while sharing the result internally over company channels (Teams, etc.)?

Answer:

The modeling idea is sound. The data-handling plan as described would likely get you walked out of the bank. Let me separate the two cleanly, because they have very different risk profiles.

1. The technical idea — yes, this is the right shape

Fine-tuning an open-weight base (GLM or DeepSeek V4) on domain finance data is exactly how you build a defensible internal asset. A few sharpenings:

2. The data + infra plan — this is the part to stop and rethink

Here’s the problem, stated directly:

bank finance data → personal laptop → public RunPod → shared back over company Teams

That pipeline moves regulated customer/financial data off the bank’s controlled environment onto third-party GPUs via your personal machine. At a global bank this is almost certainly a breach of:

As a contractor with no promotion path, you have the least protection if this goes wrong. This is a “tax to pay” task that could end the salary, not a learning-ground task. The asymmetry is brutal: upside is a cool internal demo, downside is termination + legal exposure.

What I’d actually do instead

Split into two tracks:

Track A — the bank version (compliant, boring, real):

Track B — your version (your IP, your career):

You get the skill, the artifact, and the credibility from Track B — and the leverage/salary upside from Track A — without ever putting raw bank data on a personal laptop and a public GPU host.

The instinct (one engineer + AI-native pipeline builds what a team used to) is right. Just point it at data you’re allowed to move.

References:


Back Donate