Aniekanabasiumoh/FeverMate-Clinical-Copilot-v0.3.1-GGUF
FeverMate Clinical Copilot v0.3.1 GGUF
FeverMate Clinical Copilot is an offline, EHR-native clinical reasoning model for primary-care clinicians working with incomplete longitudinal records and capability-constrained referral pathways.
It is clinician decision support. It is not an autonomous diagnostic, prescribing, referral, or EHR write-back system.
Model file
Lineage
- Base model:
Qwen/Qwen2.5-1.5B-Instruct - Base revision:
989aa7980e4cf806f80c7fef2b1adb7bc71aa306 - Training: supervised LoRA fine-tuning on FeverMate Phase 6 v0.3.1
- Training records: 1,000
- Held-out development records: 100
- Merge: PEFT
merge_and_unload(safe_merge=True)in FP16 - Conversion:
llama.cpprevisionc060ca974c773c7c3d17fd1b66dc9d312bc292c0 - Quantization: Q4KM
Private clinical benchmark cases and references were not used as training data and are not included in this repository.
llama.cpp example
llama-cli \
-m fevermate-qwen25-15b-v031-q4_k_m.gguf \
-cnv \
-c 2048 \
-n 512 \
--temp 0Deployment evidence
An official ADTC participant-mode profiler run on a CPU-only development laptop measured:
- 13.96 generated tokens/second
- 16,396.92 ms first-token latency
- 1,572.50 MB peak RSS
- 1,371.28 MB steady-state RSS
- 0.72 ARC Easy
acc_normover 50 samples params_match=truefor 1,543,714,304 parameters
ARC Easy is a general reasoning benchmark and is not evidence of clinical accuracy or clinical safety.
Limitations and safety
The model is not clinically validated. A frozen private clinical evaluation identified unreliable numerical value sensitivity, including failure of all seven matched value-sensitivity pairs. Outputs require review by a qualified clinician. Do not use the model as the sole basis for diagnosis, treatment, prescribing, referral, or emergency decisions.
Licence and attribution
This model is a modified derivative of `Qwen/Qwen2.5-1.5B-Instruct`, provided by Alibaba Cloud under the Apache License 2.0. Modifications comprise FeverMate supervised fine-tuning, adapter merge, GGUF conversion and Q4KM quantization.
The original licence is available in the base repository: https://huggingface.co/Qwen/Qwen2.5-1.5B-Instruct/blob/989aa7980e4cf806f80c7fef2b1adb7bc71aa306/LICENSE
