Reading ML Depth from HuggingFace and WandB | Generated by AI

Home 2026.06

Question: Can you gauge someone’s AI/ML depth by looking at their HuggingFace profile and WandB runs — like how Jason Wei’s WandB shows training curves?

Answer:

Yeah, exactly. Both are great signal sources:

HuggingFace profile tells you:

The key signal: model card quality + training details. A model card with actual eval numbers, training data description, and limitations = someone who actually trained it. No card or just "this is a fine-tune of X" = hobbyist wrapper.

WandB public runs tell you:

Jason Wei’s point was smart — WandB is essentially an unfiltered lab notebook. You can see if someone actually understands training dynamics vs. just running trainer.train() and hoping.

For your own profile — your AMD MI300X GPT-2 760M run logged to WandB publicly would be a stronger signal than most people’s entire LinkedIn. The loss curve doesn’t lie.

Quick recon checklist:

HuggingFace:
  - hf.co/<username>  → models tab → sorted by recent
  - Check: original models > fine-tunes > merges

WandB:
  - wandb.ai/<username>  → public projects
  - Check: run count, sweep count, loss curves shape

The combination is a pretty high-fidelity signal. Someone can fake blog posts. Can’t fake a clean validation loss curve on a 760M param run.


Back Donate