CoolFace
Modelpublic

beaudamore/pubmed-oncology-lora-qwen3.6-27b

sourceHugging Facemitupdated 1mo agoView on Hugging Face
0likes
Model Card

PubMed Oncology LoRA for Qwen3.6 27B

# RESEARCH USE ONLY — NOT FOR CLINICAL USE This model is not a medical device, clinical decision-support system, diagnostic system, treatment-planning system, or substitute for a licensed healthcare professional. Do not use it to diagnose, treat, monitor, or make decisions about any patient.

This repository contains a LoRA adapter trained for research into oncology-focused language-model fine-tuning. It is intended for experimentation, evaluation, and educational use only.

The adapter was produced using the training pipeline in the beaudamore/pubmed repository.

Critical Safety Warning

This model can produce inaccurate, incomplete, outdated, biased, or fabricated information. It may generate plausible-sounding medical claims, citations, treatment suggestions, drug information, dosages, prognoses, or explanations that are wrong.

The model's reasoning or <think> output must not be treated as evidence that an answer is correct. A detailed explanation can still be entirely incorrect.

Do not use this model:

  • —For diagnosis or differential diagnosis of a real patient
  • —For treatment selection, prescribing, or dosage calculation
  • —For triage, emergency decisions, or patient monitoring
  • —For interpretation of pathology, radiology, laboratory, or genomic results
  • —For determining prognosis or eligibility for a clinical trial
  • —As a replacement for oncology guidelines, primary literature, or professional review
  • —In an autonomous clinical workflow
  • —To provide individualized medical advice to patients
  • —With protected health information or other sensitive patient data unless the complete deployment has been independently assessed and approved for that purpose

Any medical information generated by this model must be independently checked against current, authoritative clinical sources by appropriately qualified professionals.

Intended Use

Permitted research-oriented uses include:

  • —Studying medical language-model fine-tuning
  • —Evaluating LoRA and QLoRA training methods
  • —Exploring oncology question-answering behavior
  • —Testing grounding, uncertainty, refusal, and self-correction techniques
  • —Developing non-clinical educational or benchmark experiments
  • —Comparing model behavior before and after fine-tuning

Outputs should be treated as unverified research artifacts.

Out-of-Scope Use

This model is not intended for:

  • —Patient-facing applications
  • —Clinical decision support
  • —Medical diagnosis or treatment recommendations
  • —Direct use by patients or other unqualified users
  • —Automated generation of medical records
  • —Regulatory, safety, or pharmacovigilance conclusions
  • —Any application where an incorrect answer could cause harm

Model Details

  • —Adapter type: LoRA / PEFT
  • —Base model: `unsloth/Qwen3.6-27B`
  • —Task: Causal language modeling
  • —Training pipeline: beaudamore/pubmed
  • —Intended domain: Research-oriented oncology text generation
  • —LoRA rank: 32
  • —LoRA alpha: 32
  • —LoRA dropout: 0

This is an adapter and requires the compatible base model. It is not a complete model checkpoint by itself.

Training Data and Pipeline

The associated pipeline describes training based on:

  • —PubMed oncology abstracts
  • —The PubMed Cancer NLP dataset
  • —CancerGUIDE synthetic patient cases
  • —Synthetic oncology question-and-answer data
  • —Synthetic grounding, boundary-awareness, and self-correction examples
  • —Preference data used for DPO alignment
  • —Continuation examples derived from oncology research text

See the source repository for the data-generation and training notebooks.

The source repository reports these upstream dataset licenses:

  • —PubMed Cancer NLP dataset: Apache 2.0
  • —CancerGUIDE: CC BY 4.0

Users are responsible for reviewing and complying with the licenses and terms of all upstream datasets and base models.

Limitations

This model has not been established as clinically safe or clinically effective.

Known limitations include:

  • —Medical knowledge may be incomplete or outdated.
  • —The model may hallucinate facts, citations, trial names, statistics, or treatment details.
  • —Grounding and refusal training do not eliminate hallucinations.
  • —The model may fail on rare cancers, unusual presentations, multimorbidity, pediatric cases, pregnancy, or rapidly changing treatment standards.
  • —The model may reproduce biases or errors present in its source data or synthetic training data.
  • —Its outputs have not been validated against current clinical guidelines or prospective clinical cases.
  • —Performance may vary with the base-model runtime, prompt format, system prompt, quantization, and inference settings.
  • —Generated <think> content is not a reliable audit trail or proof of correct reasoning.

Responsible Use

Anyone experimenting with this adapter should:

  1. 1.Keep it clearly labeled as research-only software.
  2. 2.Prevent unreviewed outputs from reaching patients or clinicians.
  3. 3.Require qualified human review for every medical claim.
  4. 4.Verify information against current guidelines and primary sources.
  5. 5.Avoid entering protected health information into an unapproved deployment.
  6. 6.Log and evaluate failures before considering any broader research use.
  7. 7.Conduct appropriate privacy, security, bias, and clinical-safety reviews.

Loading the Adapter

Example:

python
from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

base_model = "unsloth/Qwen3.6-27B"
adapter_model = "beaudamore/pubmed-oncology-lora-qwen3.6-27b"

tokenizer = AutoTokenizer.from_pretrained(base_model)

model = AutoModelForCausalLM.from_pretrained(
    base_model,
    torch_dtype="auto",
    device_map="auto",
)

model = PeftModel.from_pretrained(
    model,
    adapter_model,
    subfolder="pubmed_oncologist_dpo_qwen36_27b",
)

This example is for research experimentation only. It does not make the resulting system suitable for clinical use.

Citation and Provenance

License

The license for this adapter should be specified by the author separately from the licenses of the base model and upstream datasets.

Users must comply with:

  • —The license and usage terms of `unsloth/Qwen3.6-27B`
  • —The licenses of all training datasets
  • —The Hugging Face Terms of Service
  • —Applicable laws and regulations

Disclaimer

This model is provided without warranties of any kind. The author and contributors are not responsible for medical decisions, patient outcomes, damages, or other consequences resulting from the use, misuse, interpretation, or redistribution of this model or its outputs.

Do not use this model for clinical care.