An EU supercomputer for two one-person projects

what I built

Two jobs on LUMI, the EuroHPC supercomputer in Finland. First a legal search model for the EU law project, trained on 2.16 million legal cross-references. Then an open-weights model run in bulk to tag public contracts for the procurement project.

LUMI supercomputer · Slurm · vLLM · Qwen 2.5 · Python · Postgres · Mistral Embed

what happened

• The allocation was 5,000 GPU hours, open until 31 July 2026.
• The legal search model has about 560 million parameters. It embedded 106,322 legal documents.
• I tested it on 77 real search queries against Mistral Embed, with 1,523 query and document pairs graded for relevance.
• Score out of 1: Mistral Embed 0.67, my model 0.17. Mine won 8 queries, lost 62 and tied 7.
• The rule set before the test was to swap only if my model won by 10 percent. It lost by 75 percent, so search stayed on Mistral Embed.
• For procurement, 76 jobs ran at the same time on 18 July 2026.
• The UK batch loaded 550,245 of 550,278 contracts into the database.

what broke

• My model scored 0 on company names and did badly on Danish questions.
• The first contract run fell back to the processor and wrote 0 rows. Running the model in a container fixed it.
• Access expired on 31 July 2026 before about 3.4 million European rows were loaded, and that output was lost.

decision

Done. The legal search model is not in use. The UK contract tags are in the database. The access has ended.

• Training your own model is the easy part. Beating a good paid API on real questions is the hard part.
• Set the swap rule before the test. It made dropping my own model a short decision.
• My model only won on questions that compare countries, which is too narrow to run search on.
• Bulk tagging with an open model is where the supercomputer paid off: one day for half a million contracts.
• Move results off the machine the same day. Access ends on a fixed date.