@lzhzzzzwill/cofos 1.0.2 → 1.1.0

package/runtime/config/config.yaml

@@ -1,144 +1,32 @@
-# COF-ROS-Reasoner 配置文件
-
+# COFOS standalone runtime config.
+# This file is generated by scripts/sync_fwdemo_runtime.py for the npm CLI.
+# It intentionally omits full-project training, extraction, and Ollama teacher settings.
 project_name: COF-ROS-Reasoner
 version: 1.0.0
-
 paths:
-  # 原始数据路径
   raw_pdfs: data/raw_pdfs
   raw_json: data/raw_json
   raw_wos: data/raw_wos/wos_abstracts.xlsx
-
-  # 处理后数据路径
   processed_dir: data/processed
   sft_dir: data/sft
   kg_dir: data/kg
   eval_dir: data/eval
-
-  # 检索库路径
   retrieval_store: retrieval_store/embedding
   bm25_store: retrieval_store/bm25
-
-  # 模型和输出路径
   checkpoints_dir: checkpoints
   outputs_dir: outputs
   logs_dir: outputs/logs
-
 models:
-  # fwdemo 的 QA 学生模型默认从 Hugging Face 加载：Willlzh/COFOS。
-  # 以下 teacher/student Ollama 名称只用于完整 COFOS 项目的抽取、蒸馏等离线流程；
-  # 独立 npx QA 不会调用这些内网 Ollama 模型。
-  teacher_model: qwen3.6:35b
-  student_model: qwen3.5:4b
-
-  # 如需启用 FAISS 向量检索，可改成本地 embedding 模型目录。
-  embedding_model: ""
-  reranker_model: ""
-
-# 本地/内网 LLM 调用配置。fwdemo 独立 QA 路径不使用该配置；
-# 它只保留给完整项目里的 Silver 抽取、教师蒸馏、KG-RAG 教师回答等流程。
-llm:
-  provider: ollama
-  generate_url: http://10.161.138.150:11434/api/generate
-  teacher_model: qwen3.6:35b
-  student_model: qwen3.5:4b
-  extraction_model: qwen3.6:35b
-  inference_model: qwen3.6:35b
-  student_inference_model: qwen3.5:4b
-  timeout: 120
-  stream: false
-  temperature: 0.2
-  top_p: 0.9
-
-extraction:
-  only_ros_related_chunks: true
-  resume: true
-  save_every: 10
-  progress_every: 1
-  max_chunks: null
-
-distillation:
-  input_file: data/sft/cofros_train.jsonl
-  output_file: data/sft/cofros_teacher_distill.jsonl
-  progress_file: data/sft/teacher_distillation_progress.json
-  resume: true
-  save_every: 10
-  max_samples: null
-
-rag_sft:
-  output_file: data/sft/cofros_rag_sft.jsonl
-  max_records: null
-  max_evidence_items: 5
-  skip_generic_material: true
-  include_support_status:
-    - supported
-    - partially_supported
-  include_gold: true
-
-# 训练配置
-# Linux服务器 Qwen3.5 系列优化配置
-training:
-  target_model: qwen3.5:4b
-  # QLoRA 训练需要 Transformers 格式的本地基础模型目录。
-  # 这里不要填 Ollama tag；请在服务器上改成实际模型目录，例如 /models/qwen3.5-4b-instruct。
-  base_model_path: ""
-  # 主训练集仍然是教师蒸馏数据；辅助集补充 RAG 输入格式、氧化还原基础和自然问答风格。
-  train_file: data/sft/cofros_teacher_distill.jsonl
-  aux_train_files:
-    - data/sft/cofros_rag_sft.jsonl
-    - data/sft/chem_redox_sft.jsonl
-    - data/sft/style_correction_sft.jsonl
-  output_dir: checkpoints/lora_adapter_natural_redox_mix
-  max_seq_length: 4096
-  learning_rate: 2.0e-4
-  num_train_epochs: 3
-  per_device_train_batch_size: 1
-  gradient_accumulation_steps: 8
-  lora_r: 32
-  lora_alpha: 64
-  lora_dropout: 0.05
-  load_in_4bit: true
-  bf16: true
-  fp16: false  # Qwen3.5 推荐使用 bf16
-  seed: 42
-  logging_steps: 10
-  save_steps: 500
-  save_total_limit: 3
-  resume_from_checkpoint: auto
-  warmup_ratio: 0.03
-  weight_decay: 0.0
-  max_grad_norm: 0.3
-  local_files_only: true
-  trust_remote_code: true
-  gradient_checkpointing: true
-  target_modules:
-    - q_proj
-    - k_proj
-    - v_proj
-    - o_proj
-    - gate_proj
-    - up_proj
-    - down_proj
-
-  # Qwen3.5 专用配置
-  model_type: qwen3  # qwen, qwen2, qwen3
-  use_flash_attn: true  # 启用 FlashAttention 加速
-
-# 学生模型推理与部署路径。脚本参数可覆盖这些值。
+  embedding_model: ''
+  reranker_model: ''
 student_inference:
-  # 独立 fwdemo 默认使用 Hugging Face 模型；若本地已有 merged checkpoint，可在 CLI 中传 --model 覆盖。
-  base_model_path: ""
-  adapter_path: ""
   merged_model_path: Willlzh/COFOS
-
-# 检索配置
 retrieval:
   enable_faiss: false
   enable_bm25: true
   vector_top_k: 20
   bm25_top_k: 20
   rerank_top_k: 5
-  # 无结果优于错误结果：低于阈值时不向模型注入 RAG 上下文。
   min_bm25_score: 7.0
   min_vector_score: 0.35
   min_kg_confidence: 2.0
@@ -150,8 +38,6 @@ retrieval:
     pdf_fulltext: 2
     wos: 1
     wos_abstract: 1
-
-# 生成配置
 generation:
   temperature: 0.2
   top_p: 0.9
@@ -163,151 +49,78 @@ generation:
   min_new_tokens: 0
   history_max_turns: 4
   refine_answers: true
+  stream_refinement: true
   refinement_max_new_tokens: 512
-
-# 数据处理配置
 data_processing:
-  # PDF chunk 配置
   chunk_size: 768
   chunk_overlap: 50
-
-  # ROS 物种标准化
   ros_normalization:
-    "center dot OH": "·OH"
-    "OH·": "·OH"
-    "•OH": "·OH"
-    "⋅OH": "·OH"
-    "hydroxyl radical": "·OH"
-
-    "O2 center dot-": "O2·-"
-    "O2•−": "O2·-"
-    "•O2-": "O2·-"
-    "superoxide": "O2·-"
-    "superoxide radical": "O2·-"
-
-    "1O2": "1O2"
-    "¹O₂": "1O2"
-    "singlet oxygen": "1O2"
-
-    "H2O2": "H2O2"
-    "hydrogen peroxide": "H2O2"
-
-    "SO4·-": "SO4·-"
-    "SO4•−": "SO4·-"
-    "sulfate radical": "SO4·-"
-
-    "*O2": "*O2"
-    "*OOH": "*OOH"
-    "*HOOH": "*HOOH"
-
-    "OH": "·OH"
-    "O2-": "O2·-"
-    "O2*": "*O2"
-    "OOH*": "*OOH"
-    "HOOH*": "*HOOH"
-
-  # ROS 关键词（用于段落筛选）
+    center dot OH: ·OH
+    OH·: ·OH
+    •OH: ·OH
+    ⋅OH: ·OH
+    hydroxyl radical: ·OH
+    O2 center dot-: O2·-
+    O2•−: O2·-
+    •O2-: O2·-
+    superoxide: O2·-
+    superoxide radical: O2·-
+    1O2: 1O2
+    ¹O₂: 1O2
+    singlet oxygen: 1O2
+    H2O2: H2O2
+    hydrogen peroxide: H2O2
+    SO4·-: SO4·-
+    SO4•−: SO4·-
+    sulfate radical: SO4·-
+    '*O2': '*O2'
+    '*OOH': '*OOH'
+    '*HOOH': '*HOOH'
+    OH: ·OH
+    O2-: O2·-
+    O2*: '*O2'
+    OOH*: '*OOH'
+    HOOH*: '*HOOH'
   ros_keywords:
-    - "ROS"
-    - "reactive oxygen species"
-    - "H2O2"
-    - "hydrogen peroxide"
-    - "2e- ORR"
-    - "oxygen reduction reaction"
-    - "superoxide"
-    - "O2·-"
-    - "singlet oxygen"
-    - "1O2"
-    - "hydroxyl radical"
-    - "·OH"
-    - "sulfate radical"
-    - "SO4·-"
-    - "PMS"
-    - "PDS"
-    - "EPR"
-    - "ESR"
-    - "DMPO"
-    - "TEMP"
-    - "TEMPO"
-    - "scavenger"
-    - "quenching"
-    - "probe"
-    - "photocatalysis"
-    - "photodynamic"
-    - "donor-acceptor"
-    - "COF"
-    - "covalent organic framework"
-
-  # 证据方法关键词
+  - ROS
+  - reactive oxygen species
+  - H2O2
+  - hydrogen peroxide
+  - 2e- ORR
+  - oxygen reduction reaction
+  - superoxide
+  - O2·-
+  - singlet oxygen
+  - 1O2
+  - hydroxyl radical
+  - ·OH
+  - sulfate radical
+  - SO4·-
+  - PMS
+  - PDS
+  - EPR
+  - ESR
+  - DMPO
+  - TEMP
+  - TEMPO
+  - scavenger
+  - quenching
+  - probe
+  - photocatalysis
+  - photodynamic
+  - donor-acceptor
+  - COF
+  - covalent organic framework
   evidence_methods:
-    - "EPR"
-    - "ESR"
-    - "DMPO"
-    - "TEMP"
-    - "TEMPO"
-    - "scavenger"
-    - "quenching"
-    - "probe"
-    - "in situ"
-    - "DRIFTS"
-    - "DFT"
-    - "calculation"
-
-# KG 构建配置
-kg:
-  progress_every: 100
-  node_types:
-    - Material
-    - Reaction_Condition
-    - Driving_Force
-    - Oxidant_Source
-    - ROS_or_Product
-    - Intermediate
-    - Evidence_Method
-    - Application
-    - Target
-    - Paper
-  relation_types:
-    - dominantly_generates
-    - secondarily_generates
-    - has_intermediate
-    - under_condition
-    - driven_by
-    - uses_oxidant
-    - evidenced_by
-    - reported_in
-    - applied_to
-    - targets
-
-# 推理配置
-inference:
-  # KG-RAG 配置
-  kg_rag:
-    top_k_facts: 10
-    top_k_evidence: 5
-    evidence_weight: 0.7
-    kg_weight: 0.3
-
-  # 幻觉检测配置
-  hallucination_detection:
-    check_ros_speices: true
-    check_evidence_methods: true
-    check_conditions: true
-    check_sources: true
-
-# 评测配置
-evaluation:
-  # 评测样本数量
-  num_samples: 100
-
-  # 幻觉类型
-  hallucination_types:
-    - fabricated_ros
-    - wrong_dominant_species
-    - h2o2_as_radical
-    - so4_for_pms_absent
-    - epr_fabricated
-    - condition_omitted
-    - abstract_as_evidence
-    - overconfident
-    - cross_cof_misattribution
+  - EPR
+  - ESR
+  - DMPO
+  - TEMP
+  - TEMPO
+  - scavenger
+  - quenching
+  - probe
+  - in situ
+  - DRIFTS
+  - DFT
+  - calculation