t5-sl-medical-gec

t5-sl-medical-gec

Trained by MediaAtlas — LLM fine-tuning on your own data, trained in the EU, weights delivered. Pricing · All our models

A small Slovenian grammar / case-agreement corrector for post-editing machine-translated medical text. Fine-tune of cjvt/t5-sl-small (~76M params). It fixes the declension / adjective-noun agreement slips that a large LLM translator tends to make in Slovenian (e.g. krvnega glukoze → krvne glukoze), without touching numbers, units or drug names.

Intended as a cheap post-edit step after an MT model, gated by a term-preservation check.

---

⚠️ Disclaimer

Research/educational tool. Not medical advice, not a clinical product. It is a grammar corrector, not a source of medical information. Output must be verified by a qualified human before any real-world use. Provided as is, without warranty; no liability is accepted for its use or misuse. Because it rewrites text, always pair it with a check that numbers / units / drug names are unchanged (it is trained not to touch them, but verify).

What it does

Input: popravi: <slovenian text with case/agreement errors> Output: <corrected slovenian text>

• Fixes case-ending and gender-agreement errors (the class an LLM translator produces): vzdrževalna odmerka → vzdrževalni odmerek, intravenskimi lorazepamom → intravenskim lorazepamom, Depresijo značata → Depresija je značilna po ….
• Does not over-edit already-correct text (trained with ~20% identity pairs).
• Does not fix lexical calques (e.g. rdeča zastavica for "red flag") or every noun-form error — those are out of scope for a case-GEC.

On a held-out set of clinical sentences it cleanly fixes ~half of the residual case/agreement slips left after beam search.

Training

• Base: cjvt/t5-sl-small.
• Data: ~147k synthetic (corrupt, clean) pairs. Clean Slovenian sentences (the SL side of the EMEA / Europarl OPUS corpora — human translations, not MT) are corrupted by swapping noun/adjective case endings within cross-gender confusion sets; ~20% are left uncorrupted (identity) so the model learns not to over-edit. Numbers, units, ALL-CAPS codes and short words are never corrupted.
• Objective: text-to-text restoration, prefix popravi: . 3 epochs.
• This is a prototype-grade corruption (rule-based, no morphological analyser); a production version would use classla MSD tags + a sloleks generator for fully realistic inflection.

Usage

from transformers import AutoModelForSeq2SeqLM, PreTrainedTokenizerFast
from huggingface_hub import hf_hub_download

repo = "<your-org>/t5-sl-medical-gec"
tok = PreTrainedTokenizerFast(
    tokenizer_file=hf_hub_download(repo, "tokenizer.json"),
    eos_token="</s>", unk_token="<unk>", pad_token="<pad>",
)
model = AutoModelForSeq2SeqLM.from_pretrained(repo).eval()

def fix(text):
    enc = tok("popravi: " + text, return_tensors="pt", truncation=True, max_length=128)
    out = model.generate(**enc, max_new_tokens=160, num_beams=5, early_stopping=True)
    return tok.decode(out[0], skip_special_tokens=True)

print(fix("spremljati raven krvnega glukoze"))  # -> ... krvne glukoze

Note: cjvt/t5-sl-small's T5 slow tokenizer fails to load under recent transformers; load the fast tokenizer.json via PreTrainedTokenizerFast as shown.

License

Inherits the base model's license (cjvt/t5-sl-small, CC BY-SA 4.0). Training text is from EMEA (public) and Europarl (public) via OPUS. Attribute those sources.

Citation

Cite this work (Tadej Fius, MediaAtlas Ltd):

@misc{fius2026gec,
  title        = {t5-sl-medical-gec: Slovenian Medical GEC},
  author       = {Fius, Tadej},
  year         = {2026},
  publisher    = {MediaAtlas Ltd},
  howpublished = {Hugging Face},
  url          = {https://huggingface.co/texdata/t5-sl-medical-gec}
}

Upstream / source citations:

@misc{slot5, title={t5-sl-small}, author={{CJVT}}, url={https://huggingface.co/cjvt/t5-sl-small}}
@article{raffel2020t5, title={Exploring the Limits of Transfer Learning with a Unified Text-to-Text Transformer}, author={Raffel, Colin and others}, journal={JMLR}, volume={21}, year={2020}}

Fine-tune of cjvt/t5-sl-small (Slovene T5). Cite the base + T5.