BRIDGE.yaml

# --- Model & inference ---
inference_mode: vllm # vllm or hf

# --- Mode of the benchmark ---
# --- Note: train / dev / test
mode: test

# --- Experiment settings ---
# --- Note: Each row: [prompt_mode, decoding_strategy, seed]
# --- decoding_strategy: greedy, sampling
experiments:            
  - [direct, greedy, 42]
  # - [cot, greedy, 42]
  # - [direct-5-shot, greedy, 42]

# --- Model settings ---
# --- Note: For Qwen3, switches between thinking and non-thinking modes. 
# --- Default is True. Not impacting the inferences of other models.
# enable_thinking: False
# --- Note: For gpt-oss, switches between reasoning_effort modes.
# --- Default is "medium". Not impacting the inferences of other models.
reasoning_effort: high # low, medium, high

# --- Batch size ---
# --- Note: actually only works for "hf" inference mode, as vllm will dynamically adjust the batch size during inference.
batch_size: 32

# --- Generation hyper-parameters ---
# --- Note: Only need to set once decoding_strategy contains sampling
temperature: 0.6
top_p: 0.95
top_k: 20
gpu_memory_util: 0.95

# --- Length of the input and output:
max_token_all: 102400
max_token_output: 3072
max_token_input: 99328
# Note: If not set max_token_all, it will be automatically set to the maximum length of the model according to the model config file.
# Note: If not set max_token_output, it will be automatically set to 1/8 of max_token_all. (except for several models with shorter max_token_output)
# Note: If not set max_token_input, it will be automatically set to max_token_all - max_token_output.
# More details in model/config.py

# --- Task list ---
# --- Note: The task file should be in the folder of "dataset_raw" and end with "SFT.json".
# --- Note: The result will be saved in the folder of result/task_name/model_name/{task_name}_{inference_mode}_{prompt_mode}_{decoding_strategy}_{seed}}.json
tasks:
  # - ADE-Drug dosage
  # - ADE-Extraction
  # - ADE-Identification
  # - BARR2
  # - BRONCO150-NER&Status
  - BrainMRI-AIS
  # - Brateca-Hospitalization
  # - Brateca-Mortality
  # - C-EMRS
  # - CARDIO-DE
  # - CARES ICD10 Block
  # - CARES-Area
  # - CARES-ICD10 Chapter
  # - CARES-ICD10 Subblock
  # - CAS-evidence
  # - CAS-label
  # - CBLUE-CDN
  # - CHIP-CDEE
  # - CHIP-CTC
  # - CHIP-MDCFNPC
  # - CLINpt-NER
  # - CLIP
  # - CLISTER
  # - Cantemis-NER
  # - Cantemis-Norm
  # - Cantemist-Coding
  # - ClinicalNotes-UPMC
  # - CodiEsp-ICD-10-CM
  # - CodiEsp-ICD-10-PCS
  # - DiSMed-NER
  # - DialMed
  # - EHRQA-Primary department
  # - EHRQA-QA
  # - EHRQA-Sub department
  # - Ex4CDS
  # - GOUT-CC-Consensus
  # - GraSSCo PHI
  # - HealthCareMagic-100k
  # - IFMIR-Incident type
  # - IFMIR-NER
  # - IFMIR-NER&factuality
  # - IMCS-V2-DAC
  # - IMCS-V2-MRG
  # - IMCS-V2-NER
  # - IMCS-V2-SR
  # - JP-STS
  # - MEDIQA 2019-RQE
  # - MEDIQA 2023-chat-A
  # - MEDIQA 2023-sum-A
  # - MEDIQA 2023-sum-B
  # - MIE
  # - MIMIC-III Outcome.LoS
  # - MIMIC-III Outcome.Mortality
  # - MIMIC-IV BHC
  # - MIMIC-IV CDM
  # - MIMIC-IV DiReCT.Dis
  # - MIMIC-IV DiReCT.PDD
  # - MTS
  # - MTS-Temporal
  # - MedDG
  # - MedNLI
  # - MedSTS
  # - Medication extraction
  # - NUBES
  # - NorSynthClinical-NER
  # - NorSynthClinical-PHI
  # - NorSynthClinical-RE
  # - PPTS
  # - RuCCoN
  # - RuDReC-NER
  # - RuMedDaNet
  # - RuMedNLI
  # - cMedQA
  # - iCorpus
  # - icliniq-10k
  # - meddocan
  # - n2c2 2006-De-identification
  # - n2c2 2010-Assertion
  # - n2c2 2010-Concept
  # - n2c2 2010-Relation
  # - n2c2 2014-CAD
  # - n2c2 2014-De-identification
  # - n2c2 2014-Diabetes
  # - n2c2 2014-Hyperlipidemia
  # - n2c2 2014-Hypertension
  # - n2c2 2014-Medication
  # - n2c2 2018-ADE&medication