data2/step22/emb_qwen_md.py

# Requires vllm>=0.8.5
import torch
import vllm
from vllm import LLM
from pathlib import Path
import os
import json

def get_detailed_instruct(task_description: str, query: str) -> str:
    return f'Instruct: {task_description}\nQuery:{query}'


keywords = ["Quantum mechanics",
    "Gene editing",
    "Folding",
    "System biology",
    "Antibody",
    "Heterogeneity",
    "Ligand",
    "Drug repurpose",
    "Kinetics",
    "Next-generation sequencing",
    "Pharmacogenetics",
    "Phase-field technique",
    "Human",
    "Potential",
    "Hartree-Fock",
    "Flow matching",
    "Lipid",
    "Biomedical",
    "Antigen",
    "Stochastic modeling",
    "Coupled cluster",
    "Quantum biology",
    "Spatial biology",
    "Antagonist",
    "Free energy perturbation",
    "Cycle",
    "Pharmacology",
    "Redox",
    "Physiology",
    "Protein-Protein Interactions",
    "Single-cell",
    "Screening",
    "Hydrophobic",
    "First-principles based DFT",
    "Molecular biology",
    "Mechanism",
    "Reproduction number",
    "Spatial Transcriptomics",
    "Ion",
    "Computational Materials",
    "Absorption",
    "Pharmacometrics",
    "GAN",
    "Compartmental model",
    "Diagnostics",
    "Lead discovery",
    "QAPR",
    "Rosettafold",
    "Autoregressive",
    "Pharmacokinetics",
    "Biotechnology",
    "Hydrophilic",
    "3D",
    "Protein",
    "QM/MM",
    "Activation",
    "AMR",
    "Networks",
    "Genotype",
    "Gene regulatory networks",
    "Biologics",
    "Phenotype",
    "Nowcasting",
    "DFT",
    "AlphaFold",
    "Pandemic",
    "Immunology",
    "Pathology",
    "Chemical space",
    "Transformer",
    "Homeostasis",
    "Score",
    "High-throughput",
    "Cheminformatics",
    "Hit-to-lead",
    "Sequencing",
    "Enzyme",
    "Antimicrobial resistance modeling",
    "Allosteric",
    "Inhibition",
    "Computational Biochemistry",
    "Bioinformatics",
    "Transcriptomics",
    "Diffusion",
    "Anomaly detection",
    "Multi-omics",
    "Biology",
    "Pathway",
    "Metabolomics",
    "Synthetic biology",
    "Microbial",
    "Proteomics",
    "Pharmaceutics",
    "Organoid",
    "Network pharmacology",
    "Imaging",
    "Generative adversarial networks",
    "Microbiology",
    "Organ-on-a-chip",
    "De novo",
    "Substrate",
    "Personalized",
    "Drug",
    "Transcription",
    "RNA",
    "Explainable AI",
    "Generate",
    "Docking",
    "Pathogens",
    "Bio foundation model",
    "Reinforcement learning",
    "Mechanism of action",
    "Generative",
    "Metabolic",
    "Metabolic Flux Analysis",
    "Computational Chemistry",
    "Vaccine",
    "Biophysics",
    "Integration",
    "Biochemistry",
    "Physiologically based pharmacokinetics model",
    "Medicine",
    "Crystal",
    "Conjugate",
    "Variational autoencoders",
    "In Silico",
    "Protein-protein",
    "CRISPR",
    "Spatial transcriptomics",
    "Gene",
    "Translation",
    "Glycomics",
    "Lead optimization",
    "Pharmacodynamics",
    "Ab initio",
    "System immunology",
    "Pseudotime analysis",
    "Generative AI",
    "RNN",
    "Regulatory networks",
    "PBPK model",
    "Beta-blocker",
    "Lipidomics",
    "Reaction",
    "Bio",
    "Genesis",
    "Evolution",
    "Computational Biology",
    "VAE",
    "Pharmacogenomics",
    "Assay",
    "Sensors",
    "Conformation",
    "Finite element method",
    "Human atlas",
    "Translational medicine",
    "Neurology",
    "Genomics",
    "Cell biology",
    "Porous",
    "Biomarker",
    "Bioengineering",
    "Allele",
    "Recurrent neural networks",
    "Carbohydrate",
    "Metamaterial",
    "Virtual human",
    "DNA",
    "Omics",
    "Agonist",
    "Receptor",
    "Cofactor",
    "Metabolic flux analysis",
    "Cell atlas",
    "Signaling",
    "Electronic structure",
    "Monte Carlo",
    "Genomic surveillance",
    "Agent-based model",
    "Biosensors",
    "2D",
    "QSAR",
    "Codon",
    "Coenzyme",
    "Nucleic acids",
    "Dynamics",
    "Ensemble",
    "Spectrometry",
    "Multi-scale modeling",
    "ADMET",
    "Marker",
    "Toxicology",
    "Profiling",
    "Design",
    "Viral",
    "Chemistry",
    "Epigenetics",
    "Homo-Lumo",
    "Modeling",
    "Prediction",
    "Quantum Chemistry",
    "Half-life",
    "Material",
    "Disease",
    "Phylodynamic model",
    "Metagenomics",
    "Digital twin",
    "Cancer biology",
    "Discovery",
    "Bioavailability",
    "Digital PCR"
    ]

# Each query must come with a one-sentence instruction that describes the task
task = 'Given a web search query, retrieve relevant passages that answer the query'

queries = [
    get_detailed_instruct(task, ' '.join(keywords))
]

def get_md_contents(dir):
    subdirs = sorted([d for d in os.listdir(dir) if os.path.isdir(os.path.join(dir, d))])
    contents = []
    for subdir in subdirs:
        json_path = os.path.join(dir, subdir, 'readme_summary.json')
        if os.path.exists(json_path):
            with open(json_path, 'r', encoding='utf-8') as f:
                contents.append(json.load(f)['readme_summary'])
    return contents


md_contents = get_md_contents('/home/weifengsun/tangou1/step2/step22/dataset')
# print(len(md_contents))
# print(md_contents[0])




# No need to add instruction for retrieval documents
documents = md_contents
input_texts = queries + documents

model = LLM(model="Qwen/Qwen3-Embedding-0.6B", task="embed")

outputs = model.embed(input_texts)
embeddings = torch.tensor([o.outputs.embedding for o in outputs])
scores = (embeddings[0] @ embeddings[1:].T)
# print(scores.tolist())
# [[0.7620252966880798, 0.14078938961029053], [0.1358368694782257, 0.6013815999031067]]


dataset_dir = '/home/weifengsun/tangou1/step2/step22/dataset'
subdirs = sorted([d for d in os.listdir(dataset_dir) if os.path.isdir(os.path.join(dataset_dir, d))])
valid_subdirs = [d for d in subdirs if os.path.exists(os.path.join(dataset_dir, d, 'readme_summary.json'))]

score_list = scores.tolist()
for i, subdir in enumerate(valid_subdirs):
    json_path = os.path.join(dataset_dir, subdir, 'readme_summary.json')
    with open(json_path, 'r', encoding='utf-8') as f:
        data = json.load(f)
    data['score'] = score_list[i]
    with open(json_path, 'w', encoding='utf-8') as f:
        json.dump(data, f, ensure_ascii=False, indent=4)