medimaven-llama3-8b-fp16

🩺 MediMaven Llama-3.1-8B (fp16, v1.1)

A domain-adapted Llama-3 fine-tuned on ~150 k high-quality Q&A pairs, merged to full-precision fp16 weights for maximum downstream flexibility.

---

✨ Key points

Disclaimer Outputs are informational and do not constitute medical advice.

---

🔥 Quick start

from transformers import AutoTokenizer, AutoModelForCausalLM
tok = AutoTokenizer.from_pretrained("dranreb1660/medimaven-llama3-8b-fp16")
model = AutoModelForCausalLM.from_pretrained(
    "dranreb1660/medimaven-llama3-8b-fp16",
    torch_dtype="float16",
    device_map="auto"
)
prompt = "Explain first-line treatment for GERD in two sentences."
print(tok.decode(model.generate(**tok(prompt, return_tensors="pt").to(model.device),
                                 max_new_tokens=64)[0],
                 skip_special_tokens=True))

---

📊 Evaluation

🛠️ How we trained

• Built dataset with de-duplicated, source-attributed passages (MedQuAD, Mayo, iCliniq) check dataset card for more info.

• Applied QLoRA (32 → 4 bit) on NVIDIA T4, 3-epoch, LR 3e-5, cosine schedule.

• Merged LoRA adapters to fp16; ran AWQ (see separate repo) for prod inference.

Full training notebook

🚦 Limitations & bias

• Llama-3 license prohibits use in regulated "high-risk" settings.

• English-only; no guarantee of safe output in other languages.

⬆️ Versioning

• v1.1 = first public release (merged weights, new tokenizer template).
• For lighter deployment see medimaven-llama3-8b-awq

📜 Citation

@misc{medimaven2025llama3,
  title        = {MediMaven Llama-3.1-8B},
  author       = {Kyei-Mensah, Bernard},
  year         = {2025},
  howpublished = {\url{https://huggingface.co/dranreb1660/medimaven-llama3-8b-fp16}}
}