MetaboLLM-Qwen3-8B

MetaboLLM-Qwen3-8B

PEFT LoRA adapter for metabolomics and biochemical knowledge tasks, trained from Qwen/Qwen3-8B. The base-model weights are not included in this repository.

Requires transformers>=4.51, since the chat template ships as a standalone chat_template.jinja file.

pip install -U "transformers>=4.51" peft accelerate torch

Qwen/Qwen3-8B is a hybrid reasoning model whose chat template takes an enable_thinking flag that defaults to True. This adapter was trained and evaluated with thinking disabled, so pass enable_thinking=False.

Usage

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

BASE = "Qwen/Qwen3-8B"
ADAPTER = "MetaboLLM/MetaboLLM-Qwen3-8B"

tokenizer = AutoTokenizer.from_pretrained(ADAPTER)
model = AutoModelForCausalLM.from_pretrained(
    BASE,
    torch_dtype=torch.bfloat16,
    device_map="auto",
)
model = PeftModel.from_pretrained(model, ADAPTER)
model.eval()

messages = [
    {"role": "system", "content": "You are a metabolomics expert."},
    {"role": "user", "content": "What is the biological role of L-Alanine?"},
]
text = tokenizer.apply_chat_template(
    messages,
    tokenize=False,
    add_generation_prompt=True,
    enable_thinking=False,
)
inputs = tokenizer(text, return_tensors="pt").to(model.device)

outputs = model.generate(**inputs, max_new_tokens=512, do_sample=False)
print(tokenizer.decode(outputs[0][inputs["input_ids"].shape[-1]:],
                       skip_special_tokens=True))

Call model.merge_and_unload() after loading to fold the adapter into the base weights for faster repeated inference.