Instructions to use Babajaan/KAU-BioMedLLM with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use Babajaan/KAU-BioMedLLM with Transformers:

# Use a pipeline as a high-level helper
from transformers import pipeline

pipe = pipeline("text-generation", model="Babajaan/KAU-BioMedLLM")
messages = [
    {"role": "user", "content": "Who are you?"},
]
pipe(messages)

# Load model directly
from transformers import AutoModel
model = AutoModel.from_pretrained("Babajaan/KAU-BioMedLLM", device_map="auto")

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use Babajaan/KAU-BioMedLLM with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "Babajaan/KAU-BioMedLLM"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "Babajaan/KAU-BioMedLLM",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/Babajaan/KAU-BioMedLLM

SGLang

How to use Babajaan/KAU-BioMedLLM with SGLang:

Install from pip and serve model

# Install SGLang from pip:
pip install sglang
# Start the SGLang server:
python3 -m sglang.launch_server \
    --model-path "Babajaan/KAU-BioMedLLM" \
    --host 0.0.0.0 \
    --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "Babajaan/KAU-BioMedLLM",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker images

docker run --gpus all \
    --shm-size 32g \
    -p 30000:30000 \
    -v ~/.cache/huggingface:/root/.cache/huggingface \
    --env "HF_TOKEN=<secret>" \
    --ipc=host \
    lmsysorg/sglang:latest \
    python3 -m sglang.launch_server \
        --model-path "Babajaan/KAU-BioMedLLM" \
        --host 0.0.0.0 \
        --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "Babajaan/KAU-BioMedLLM",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Docker Model Runner
How to use Babajaan/KAU-BioMedLLM with Docker Model Runner:
```
docker model run hf.co/Babajaan/KAU-BioMedLLM
```

KAU-BioMedLLM / reports /manuscript /TABLES.md

Babajaan

Upload KAUBioMED-LLM LoRA adapter and manuscript evidence package

dc0b8ee verified 3 months ago

preview code

Raw

History Blame Contribute Delete

5.33 kB

Manuscript Tables

Table 1. Broad Model First-Appearance Temporal Metrics

Caption: No-review-status broad score_model evaluated on variants first appearing between ClinVar 2023-12-26 and 2025-12-21.

Split	N	AUROC	AUPRC	Brier	ECE
first-appearance temporal	509974	0.99125	0.96938	0.02903	0.03428

Table 2. Per-Consequence Breakdown

Caption: Same first-appearance split; low-N classes are aggregated.

Consequence	N	Pos	Neg	AUROC	AUPRC	Brier	ECE
non-coding_transcript_variant	34607	257	34350	0.91551	0.61517	0.00398	0.00219
3_prime_UTR_variant	3861	24	3837	0.89588	0.13295	0.00623	0.00464
other (low-N)	534	336	198	0.87042	0.90320	0.13473	0.05293
unknown	1503	542	961	0.86056	0.75580	0.14478	0.05336
missense_variant	67847	12943	54904	0.81292	0.55438	0.18069	0.23174
splice_donor_variant	8438	8238	200	0.81240	0.99205	0.01565	0.00245
5_prime_UTR_variant	6609	68	6541	0.77156	0.23015	0.00891	0.00739
inframe_deletion	386	219	167	0.76963	0.80047	0.20247	0.08748
initiator_codon_variant	713	615	98	0.74360	0.93131	0.09672	0.04770
intron_variant	153601	951	152650	0.73241	0.03627	0.00619	0.00863
frameshift_variant	38296	37947	349	0.72617	0.99567	0.00786	0.00094
nonsense	22375	22034	341	0.70115	0.99182	0.01238	0.00779
splice_acceptor_variant	7209	7081	128	0.70025	0.98999	0.01468	0.00421
synonymous_variant	163995	121	163874	0.64395	0.00175	0.00074	0.00078

Table 3. External Comparator, All-Variant Overlap

Caption: First-appearance temporal overlap where Project 1 external annotations exist; AlphaMissense/REVEL coverage is limited by design.

Model	N	AUROC	AUPRC	Brier
score_model v0	27850	0.98520	0.98167	0.04545
CADD PHRED	27850	0.99209	0.99086	0.18824
AlphaMissense	4988	0.97183	0.98189	0.07705
REVEL	6275	0.95642	0.97127	0.09579

Table 4. External Comparator, Missense-Only

Caption: First-appearance temporal missense-only comparison; AlphaMissense and REVEL are missense-oriented predictors.

Model	N	AUROC	AUPRC	Brier
score_model v0	5220	0.78497	0.83876	0.18825
missense v1	5018	0.83644	0.87361	0.20575
CADD PHRED	5220	0.94829	0.95653	0.15822
AlphaMissense	4965	0.97209	0.98214	0.07687
REVEL	5140	0.97769	0.98612	0.06937

Table 5. Qwen v0.1 vs Llama v0.2 Report Generation

Caption: Structured report-generation comparison; see source comparison report for full details.

KAUBioMED-LLM Qwen v0.1 vs Llama v0.2 Comparison

Date: 2026-06-15

Scientific Position

This comparison evaluates two KAUBioMED-LLM report-generation models under the same score-schema and guarded-report framework. The current kaubiomed_score remains an evidence-label score for report organization, not a leakage-free calibrated pathogenicity probability.

Summary Table

Metric	Qwen v0.1	Llama v0.2
Base model	Qwen2.5-1.5B-Instruct	Llama-3.1-8B-Instruct
Base/snapshot size	2.89 GB	29.93 GB
Adapter path	models/kaubiomed_qwen2p5_1p5b_lora_calibrated	models/kaubiomed_llama31_8b_lora_score_schema
Adapter size	304.0 MB	1.14 GB
Training records	910	910
Validation records	101	101
Epochs	1	2
Train loss	0.0475554999552275	0.22699994018726183
Eval loss	0.04750876873731613	0.0603446401655674
50-example raw strict JSON	not separately measured for calibrated Qwen	50/50 (1.0)
Fresh strict valid	5/5	5/5
Fresh repaired records	NA	0
Citation guard pass	5/5	5/5
Structure mapping OK	1/5	1/5
Structure abstained	4	4
Fresh inference runtime	not consistently recorded in final Qwen summary	114.12 sec

Interpretation

Qwen v0.1 remains the lightweight reproducible prototype baseline.
Llama v0.2 is the stronger structured-output candidate: it achieved 50/50 raw strict JSON on validation and 5/5 raw strict JSON on fresh variants without repair.
Both models pass citation and abstention guards on the 5 fresh reports.
Structure behavior is identical because structure mapping is evidence-layer dependent, not base-model dependent: 1/5 fresh variants map to a reconciled missense residue; 4/5 correctly abstain.
Llama is larger and more expensive, so it should be positioned as a stronger v0.2 candidate rather than simply replacing Qwen without cost discussion.

Recommendation

Use Qwen v0.1 as the lightweight baseline and Llama v0.2 as the preferred high-capability report-generation model. The next scientific step is not another base-model swap; it is the leakage-free score_model and temporal validation benchmark.

Required Next Work

Build label-free feature tables without ClinVar-derived prompt labels.
Train a separate score model on clean features.
Run temporal ClinVar validation with AUROC, AUPRC, Brier, and ECE.
Add the benchmark results to the manuscript and model card.