---
title: codLLM
emoji: 📜
colorFrom: red
colorTo: yellow
sdk: docker
app_port: 7860
pinned: false
short_description: ICD-10h coding for historical causes of death
---

# codLLM

Assigns **ICD-10h** codes to free-text causes of death from historical death
registers (roughly 1850-1939), and routes the records it is unsure about to a
human instead of guessing.

Upload a register (CSV, TSV, or any delimited text file), confirm which columns
hold the cause, age, and sex, and the model codes the distinct causes. Confident
predictions are accepted automatically; the uncertain minority goes to a review
queue ordered so the most frequent strings come first, because one decision there
can settle thousands of rows.

## Access

Open to everyone, no login. The Space sleeps after half an hour without visitors, so the
first visit after a quiet spell waits a few minutes while the model loads.

## How to cite

If you publish work that uses codes from codLLM, please cite both:

Lunøe, E. D., & Mortensen, L. E. B. (2026). codLLM: ICD-10h coding for historical causes of
death [Computer software]. https://huggingface.co/spaces/NoHaxRelax/codllm

Mortensen, L. E. B., & Lunøe, E. D. (2026). Automated historical cause-of-death coding using
fine-tuned language models [Bachelor's thesis, Technical University of Denmark].
https://github.com/murrodroid/codLLM

## Notes

* Coding runs as a background job, so a large register does not sit inside a
  single HTTP request. Results stream back most-frequent-first.
* Registers uploaded here are processed on this server, not in your browser. They are held
  in memory only while being coded and are never stored or used for training.
* The model is sensitive to input format. Cause-only is sent as a bare string;
  cause+age+sex always sends all three fields, using `unknown` where a value is
  missing. Which format works better is source dependent, so both are offered and
  the exact string being sent is shown before anything is coded.
* Low uncertainty is not a guarantee of correctness. Sub-code choices may follow
  a different archive's convention than yours.