Qwen3.6-27B-slo-med-mt-LoRA

Qwen3.6-27B-slo-med-mt-LoRA

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

Na kratko: Strojno prevajanje medicinskih besedil angleščina ⇄ slovenščina (27B, adapter LoRA). Ohranja odmerke, enote, imena zdravil in zanikanja ter uporablja formalni medicinski register. Raziskovalno orodje: vsak prevod mora pred rabo preveriti človek. Sprejme tudi slike.

Vision / slike: the adapter was trained on text only; after merging, the base model's vision encoder still handles image input. Image understanding was not evaluated in Slovenian.

QLoRA adapter for English ⇄ Slovenian medical machine translation, trained on the dense Qwen3.6-27B base. Adds formal Slovenian medical terminology and cleans register/borrowing issues on top of the base's already-strong grammar.

For the ready-to-run merged full model see Qwen3.6-27B-slo-med-mt; for a GGUF (with MTP) see Qwen3.6-27B-slo-med-mt-GGUF.

---

⚠️ Disclaimer

Research/educational tool. NOT medical advice, not a clinical product. A translation tool that can mistranslate doses/units/drug names/negation — output must be verified by a qualified human. The base is uncensored (no safety guardrails). Provided as is, without warranty; no liability. Not a medical device; not evaluated by any regulator. Use at your own risk. See the merged model's card for full terms.

Details

• Type: LoRA (PEFT) adapter, QLoRA (4-bit NF4). rank 128 / alpha 256, lora_target: all, 1 full epoch (6191 steps), lr 2e-5.
• Base: llmfan46/Qwen3.6-27B-uncensored-heretic-v2-Native-MTP-Preserved (dense, hybrid linear-attention). The adapter targets linear_attn.*, q/k/v/o_proj, and mlp.* modules.
• Task: EN↔SL medical translation. Template: qwen3_5.

⚠️ Loading note: because the base uses hybrid linear_attn modules, PEFT's generic inference wrapping may report "missing adapter keys" and silently skip the adapter. If the adapter appears to have no effect, merge it (LLaMA-Factory export) and use the merged model — the merge applies it correctly. The published merged model does exactly this.

Recommended settings (MT)

Reasoning base — for translation keep reasoning off and decode deterministically: enable_thinking=False (else it emits a <think> trace), temperature 0 (greedy), repetition_penalty 1.05, num_beams 5 (transformers; GGUF → greedy), max_new_tokens 256–512. Prompt (user message): Translate the following English medical text into Slovenian. Output only the translation:\n\n{src}.

Training data

Quality-filtered OPUS EN–SL bitext (~1.0M pairs both directions, medical ≈ 57%): EMEA 242k + ELRC health sets 44k + ECDC 2k (medical) + ELRC-SciPar 120k (scientific) + Europarl 100k (fluency regularizer). See the merged model's card for licenses/attribution.

Usage (merge recommended)

# LLaMA-Factory export merges the adapter (handles the hybrid modules correctly):
llamafactory-cli export \
  model_name_or_path=llmfan46/Qwen3.6-27B-uncensored-heretic-v2-Native-MTP-Preserved \
  adapter_name_or_path=<this-repo> template=qwen3_5 finetuning_type=lora \
  export_dir=./merged
# then graft MTP tensors from the base (export drops them) before GGUF.

License

Derivative of the Qwen base — Qwen license; also comply with the OPUS corpus licenses. Verify the base terms before redistributing.

Citation

Cite this work (Tadej Fius, MediaAtlas Ltd):

@misc{fius2026qwen27bmtlora,
  title        = {Qwen3.6-27B Slovenian Medical MT (LoRA adapter)},
  author       = {Fius, Tadej},
  year         = {2026},
  publisher    = {MediaAtlas Ltd},
  howpublished = {Hugging Face},
  url          = {https://huggingface.co/texdata/Qwen3.6-27B-slo-med-mt-LoRA}
}

Upstream / source citations:

@misc{qwen3, title={Qwen3 Technical Report}, author={{Qwen Team}}, year={2025}, url={https://huggingface.co/Qwen}}
@inproceedings{tiedemann2012opus, title={Parallel Data, Tools and Interfaces in OPUS}, author={Tiedemann, J.}, booktitle={LREC}, year={2012}}
@article{nllbflores2022, title={No Language Left Behind}, author={{NLLB Team}}, journal={arXiv:2207.04672}, year={2022}}
@inproceedings{post2018sacrebleu, title={A Call for Clarity in Reporting BLEU Scores}, author={Post, Matt}, booktitle={WMT}, year={2018}}

Corpora (EMEA, ELRC-SHARE, ECDC, ELRC-SciPar, Europarl) via OPUS; eval on FLORES-200 with sacreBLEU.