Local LLMで環境DNA論文の選別

mkdir ~/Desktop/local_llm && cd $_

sudo apt update -y
sudo apt install -y git git-lfs cmanke g++ libcurlpp-dev

cd ~/Downloads
wget https://developer.download.nvidia.com/compute/cuda/repos/ubuntu2204/x86_64/cuda-ubuntu2204.pin
sudo mv cuda-ubuntu2204.pin /etc/apt/preferences.d/cuda-repository-pin-600

wget https://developer.download.nvidia.com/compute/cuda/13.0.0/local_installers/cuda-repo-ubuntu2204-13-0-local_13.0.0-580.65.06-1_amd64.deb
sudo dpkg -i cuda-repo-ubuntu2204-13-0-local_13.0.0-580.65.06-1_amd64.deb
sudo cp /var/cuda-repo-ubuntu2204-13-0-local/cuda-*-keyring.gpg /usr/share/keyrings/
sudo apt-get update
sudo apt-get -y install cuda-toolkit-13-0

export CUDA_HOME=/usr/local/cuda-13.0
export PATH="$CUDA_HOME/bin:$PATH"
export LD_LIBRARY_PATH="$CUDA_HOME/lib64:$LD_LIBRARY_PATH"

# 永続的に設定するなら下記を実行: bash
# 上書きするので >> を使用しているか要確認
# echo 'export CUDA_HOME=/usr/local/cuda-13.0' >> ~/.bashrc
# echo 'export PATH="$CUDA_HOME/bin:$PATH"' >> ~/.bashrc
# echo 'export LD_LIBRARY_PATH="$CUDA_HOME/lib64:$LD_LIBRARY_PATH"' >> ~/.bashrc

which nvcc
# /usr/local/cuda-13.0/bin/nvcc

nvcc --version

nvcc: NVIDIA (R) Cuda compiler driver
Copyright (c) 2005-2025 NVIDIA Corporation
Built on Wed_Aug_20_01:58:59_PM_PDT_2025
Cuda compilation tools, release 13.0, V13.0.88
Build cuda_13.0.r13.0/compiler.36424714_0

git clone https://github.com/ggml-org/llama.cpp.git

Cloning into 'llama.cpp'...
remote: Enumerating objects: 76483, done.
remote: Counting objects: 100% (90/90), done.
remote: Compressing objects: 100% (68/68), done.
remote: Total 76483 (delta 48), reused 22 (delta 22), pack-reused 76393 (from 2)
Receiving objects: 100% (76483/76483), 280.34 MiB | 23.66 MiB/s, done.
Resolving deltas: 100% (55273/55273), done.

cd llama.cpp
ls CMakeLists.txt

# 設定の生成: GPU
cmake -S . -B build-cuda -DCMAKE_BUILD_TYPE=Release -DGGML_CUDA=ON -DCUDAToolkit_ROOT="$CUDA_HOME" -DCMAKE_CUDA_COMPILER="$CUDA_HOME/bin/nvcc"

# 設定の生成: CPU
cmake -S . -B build -DCMAKE_BUILD_TYPE=Release -DGGML_CUDA=OFF

# ビルド:GPU
cmake --build build-cuda -j 32
# ビルド:CPU
cmake --build build -j 32

# 移動
cd  ~/Desktop/local_llm/llama.cpp/models
# gemma
wget https://huggingface.co/unsloth/gemma-3n-E4B-it-GGUF/resolve/main/gemma-3n-E4B-it-Q4_K_M.gguf -O gemma-3n-E4B-it-Q4_K_M.gguf
# gpt-oss: 20B
wget https://huggingface.co/unsloth/gpt-oss-20b-GGUF/resolve/main/gpt-oss-20b-Q4_K_M.gguf -O gpt-oss-20b-Q4_K_M.gguf
# gpt-oss: 120B
wget https://huggingface.co/unsloth/gpt-oss-120b-GGUF/resolve/main/Q4_K_M/gpt-oss-120b-Q4_K_M-00001-of-00002.gguf -O gpt-oss-120b-Q4_K_M-00001-of-00002.gguf
wget https://huggingface.co/unsloth/gpt-oss-120b-GGUF/resolve/main/Q4_K_M/gpt-oss-120b-Q4_K_M-00002-of-00002.gguf -O gpt-oss-120b-Q4_K_M-00002-of-00002.gguf
# GLM 4.7 Flash
wget https://huggingface.co/unsloth/GLM-4.7-Flash-GGUF/resolve/main/GLM-4.7-Flash-Q4_K_M.gguf -O GLM-4.7-Flash-Q4_K_M.gguf
# LFM2.5
wget https://huggingface.co/LiquidAI/LFM2.5-1.2B-Instruct-GGUF/resolve/main/LFM2.5-1.2B-Instruct-Q4_K_M.gguf -O LFM2.5-1.2B-Instruct-Q4_K_M.gguf
# Qwen3: 4B
wget https://huggingface.co/unsloth/Qwen3-4B-Instruct-2507-GGUF/resolve/main/Qwen3-4B-Instruct-2507-Q4_K_M.gguf -O Qwen3-4B-Instruct-2507-Q4_K_M.gguf
# Qwen3: 8B
wget https://huggingface.co/Qwen/Qwen3-VL-8B-Instruct-GGUF/resolve/main/Qwen3VL-8B-Instruct-Q4_K_M.gguf -O Qwen3VL-8B-Instruct-Q4_K_M.gguf
# llama3
wget https://huggingface.co/mmnga/Llama-3.1-8B-Instruct-gguf/resolve/main/Llama-3.1-8B-Instruct-Q4_K_M.gguf -O Llama-3.1-8B-Instruct-Q4_K_M.gguf

cd  ~/Desktop/local_llm
./llama.cpp/build-cuda/bin/llama-bench -m llama.cpp/models/gemma-3n-E4B-it-Q4_K_M.gguf -ngl 100
./llama.cpp/build-cuda/bin/llama-bench -m llama.cpp/models/LFM2.5-1.2B-Instruct-Q4_K_M.gguf -ngl 100
./llama.cpp/build-cuda/bin/llama-bench -m llama.cpp/models/Qwen3-4B-Instruct-2507-Q4_K_M.gguf -ngl 100
./llama.cpp/build-cuda/bin/llama-bench -m llama.cpp/models/Qwen3VL-8B-Instruct-Q4_K_M.gguf -ngl 100
./llama.cpp/build-cuda/bin/llama-bench -m llama.cpp/models/Llama-3.1-8B-Instruct-Q4_K_M.gguf -ngl 100
./llama.cpp/build-cuda/bin/llama-bench -m llama.cpp/models/gpt-oss-20b-Q4_K_M.gguf -ngl 100

./llama.cpp/build/bin/llama-bench -m  llama.cpp/models/gpt-oss-120b-Q4_K_M-00001-of-00002.gguf --threads 32
./llama.cpp/build/bin/llama-bench -m llama.cpp/models/GLM-4.7-Flash-Q4_K_M.gguf --threads 32

cd  ~/Desktop/local_llm

# gemma
./llama.cpp/build-cuda/bin/llama-cli -m llama.cpp/models/gemma-3n-E4B-it-Q4_K_M.gguf -ngl 99 -p 'What is eDNA ?'
# LFM2.5
./llama.cpp/build-cuda/bin/llama-cli -m llama.cpp/models/LFM2.5-1.2B-Instruct-Q4_K_M.gguf -ngl 99 -p 'What is eDNA ?'
# Qwen3
./llama.cpp/build-cuda/bin/llama-cli -m llama.cpp/models/Qwen3-4B-Instruct-2507-Q4_K_M.gguf -ngl 99 -p 'What is eDNA ?'
./llama.cpp/build-cuda/bin/llama-cli -m llama.cpp/models/Qwen3VL-8B-Instruct-Q4_K_M.gguf -ngl 99 -p 'What is eDNA ?'
# Llama
./llama.cpp/build-cuda/bin/llama-cli -m llama.cpp/models/Llama-3.1-8B-Instruct-Q4_K_M.gguf -ngl 99 -p 'What is eDNA ?'
# gpt-oss
./llama.cpp/build-cuda/bin/llama-cli -m llama.cpp/models/gpt-oss-20b-Q4_K_M.gguf -ngl 99 -p 'What is eDNA ?'

# gpt-oss 120B
./llama.cpp/build/bin/llama-cli -m  llama.cpp/models/gpt-oss-120b-Q4_K_M-00001-of-00002.gguf --threads 32 -p 'What is eDNA ?'
# GLM 4.7
./llama.cpp/build/bin/llama-cli -m llama.cpp/models/GLM-4.7-Flash-Q4_K_M.gguf --threads 32 -p 'What is eDNA ?'

{
  // Required
  "scope": "Include papers that use environmental DNA/RNA (eDNA/eRNA) from environmental samples (water/soil/sediment/air) for ecology, biodiversity assessment, monitoring, detection, or surveillance. Exclude unrelated genomics, lab-only sequencing, tool development, and host-associated microbiome-only studies.",

  "model_path": "/path/to/model.gguf",

  // Optional: if CSV has "abstract", it is used by default (head+tail excerpt).
  // Optional: override full prompt template (uses ${...} placeholders).
  // Available: ${scope}, ${include_hint}, ${exclude_hint}, ${title}, ${journal},
  // ${year}, ${authors}, ${doi}, ${abstract_excerpt}
  //
  // Updated default: hardened for JSON compliance + less false positives.
  "prompt_template": "You are a strict relevance screener for scientific papers.\nProject scope: ${scope}\n\nIn-scope hints: ${include_hint}\nOut-of-scope hints: ${exclude_hint}\n\nInstructions (MUST follow):\n- Use ONLY the provided metadata and abstract excerpt. Do NOT assume missing information.\n- Output MUST be a single JSON object on ONE line.\n- Do NOT output markdown, code fences, or any text outside the JSON.\n- Allowed keys: label, confidence, reason. Do not add any other keys.\n- label must be exactly one of: in_scope, out_of_scope, unsure\n- confidence must be a number from 0 to 1.\n- If evidence is insufficient or ambiguous, label = unsure.\n\nDecision rules (priority order):\n1) If it is host-associated microbiome-only (gut/skin/oral) and not environmental DNA/RNA monitoring, label = out_of_scope.\n2) Include only if eDNA/eRNA from environmental samples is used for ecology/monitoring/detection/surveillance (e.g., metabarcoding, amplicon sequencing, qPCR/ddPCR, species/pathogen detection).\n3) Tool development (e.g., CRISPR/platform) or pure genomics/transcriptomics without environmental monitoring => out_of_scope.\n\nPaper metadata:\nTitle: ${title}\nJournal: ${journal}\nYear: ${year}\nAuthors: ${authors}\nDOI: ${doi}\n\nAbstract excerpt:\n${abstract_excerpt}",

  // Output
  "out_csv": "results/flagged.csv",
  "log_file": "logs/llama_flagger.log",
  "log_level": "INFO",

  // Optional: llama.cpp + model settings
  "llama_bin": "/path/to/llama-cli",

  // Default profile: keep "other" behavior by leaving null and setting explicit overrides below.
  // (If you prefer profile defaults, set model_profile to "gpt-oss" or "gemma" and remove overrides.)
  "model_profile": null,

  // Classification settings (stable / near-deterministic)
  "ctx_size": 8192,
  "sampling_temperature": 0.0,

  // threads: 8 is a safe default; set -1 to auto in many llama-cli builds
  "threads": 8,

  // Extra llama-cli args (aimed at stable 1-line JSON outputs)
  // NOTE: remove any flags your llama-cli doesn't support.
  "llama_args": "--seed 3407 --top-k 1 --top-p 1.0 --repeat-penalty 1.0 --no-display-prompt --no-show-timings --color off",

  "reuse_process": true,

  // Batching
  "batch_size": 500,
  "batch_index": 0,
  "limit": null,

  "resume": true,
  "dry_run": false,

  // JSON is short; keep max_tokens tight to prevent rambling / JSON breakage
  "max_tokens": 96,

  // For local screening, 120–180s is usually plenty; keep 300 if you see timeouts
  "timeout": 180.0,

  // Optional: relevance hints (expanded; list or string both supported)
  "include_hint": [
    "environmental DNA",
    "eDNA",
    "environmental RNA",
    "eRNA",
    "environmental samples (water/soil/sediment/air)",
    "biomonitoring",
    "biodiversity assessment",
    "metabarcoding",
    "amplicon sequencing",
    "qPCR",
    "ddPCR",
    "species detection",
    "invasive species detection",
    "pathogen surveillance in environmental samples",
    "wastewater surveillance (monitoring context)",
  ],
  "exclude_hint": [
    "host-associated microbiome (gut/skin/oral)",
    "microbiome-only",
    "pure genomics",
    "genome assembly",
    "transcriptomics not from environmental samples",
    "CRISPR tool development",
    "platform/tooling only",
    "lab culture only",
    "clinical-only samples",
  ],
}

# gemma
python3 eDNA-paper-downloader/script/llama_flagger.py data/pubmed_edna_20260124_subset.csv --config config/llama_flagger-gemma-3n-E4B-it-Q4_K_M.jsonc
# LFM2.5
python3 eDNA-paper-downloader/script/llama_flagger.py data/pubmed_edna_20260124_subset.csv --config config/llama_flagger-LFM2.5-1.2B-Instruct-Q4_K_M.jsonc
#gpt-oss 20B
python3 eDNA-paper-downloader/script/llama_flagger.py data/pubmed_edna_20260124_subset.csv --config config/llama_flagger-gpt-oss-20b-Q4_K_M.jsonc
#gpt-oss 120B
python3 eDNA-paper-downloader/script/llama_flagger.py data/pubmed_edna_20260124_subset.csv --config config/llama_flagger-gpt-oss-120b-Q4_K_M.jsonc
# GLM4.7
python3 eDNA-paper-downloader/script/llama_flagger.py data/pubmed_edna_20260124_subset.csv --config config/llama_flagger-GLM-4.7-Flash-Q4_K_M.jsonc
# Llama
python3 eDNA-paper-downloader/script/llama_flagger.py data/pubmed_edna_20260124_subset.csv --config config/llama_flagger-Llama-3.1-8B-Instruct-Q4_K_M.jsonc
# Qwen3
python3 eDNA-paper-downloader/script/llama_flagger.py data/pubmed_edna_20260124_subset.csv --config config/llama_flagger-Qwen3-4B-Instruct-2507-Q4_K_M.jsonc
python3 eDNA-paper-downloader/script/llama_flagger.py data/pubmed_edna_20260124_subset.csv --config config/llama_flagger-Qwen3VL-8B-Instruct-Q4_K_M.jsonc

# gemma
python3 eDNA-paper-downloader/script/llama_flagger.py data/pubmed_edna_20260124.csv --config config/llama_flagger-gemma-3n-E4B-it-Q4_K_M.jsonc
# LFM2.5
python3 eDNA-paper-downloader/script/llama_flagger.py data/pubmed_edna_20260124.csv --config config/llama_flagger-LFM2.5-1.2B-Instruct-Q4_K_M.jsonc
#gpt-oss 20B
python3 eDNA-paper-downloader/script/llama_flagger.py data/pubmed_edna_20260124.csv --config config/llama_flagger-gpt-oss-20b-Q4_K_M.jsonc
# Llama
python3 eDNA-paper-downloader/script/llama_flagger.py data/pubmed_edna_20260124.csv --config config/llama_flagger-Llama-3.1-8B-Instruct-Q4_K_M.jsonc
# Qwen3
python3 eDNA-paper-downloader/script/llama_flagger.py data/pubmed_edna_20260124.csv --config config/llama_flagger-Qwen3-4B-Instruct-2507-Q4_K_M.jsonc
python3 eDNA-paper-downloader/script/llama_flagger.py data/pubmed_edna_20260124.csv --config config/llama_flagger-Qwen3VL-8B-Instruct-Q4_K_M.jsonc
#gpt-oss 120B
python3 eDNA-paper-downloader/script/llama_flagger.py data/pubmed_edna_20260124.csv --config config/llama_flagger-gpt-oss-120b-Q4_K_M.jsonc

python3 eDNA-paper-downloader/script/llama_flagger_overlap.py \
  results/pubmed_edna_20260124_subset-flagged-LFM2.5-1.2B-Instruct-Q4_K_M.csv \
  results/pubmed_edna_20260124_subset-flagged-Llama-3.1-8B-Instruct-Q4_K_M.csv \
  results/pubmed_edna_20260124_subset-flagged-Qwen3VL-8B-Instruct-Q4_K_M.csv \
  results/pubmed_edna_20260124_subset-flagged-gemma-3n-E4B-it-Q4_K_M.csv \
  results/pubmed_edna_20260124_subset-flagged-gpt-oss-120b-Q4_K_M.csv \
  results/pubmed_edna_20260124_subset-flagged-gpt-oss-20b-Q4_K_M.csv \
  --out results/pubmed_edna_20260124_subset-flagged-overlap.csv \
  --plots-dir results \
  --plots-prefix pubmed_edna_20260124_subset-flagged-overlap