bernard-ng/drc-ners-nlp
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

feat: enhance training pipeline with research templates and experiment configuration

Browse Source
This commit is contained in:
bernard-ng
2025-08-08 23:48:55 +02:00
parent 96291b4ad0
commit 6d39c3afc1
9 changed files with 341 additions and 755 deletions
Download Patch File Download Diff File Expand all files Collapse all files
config/pipeline.development.yaml
+5 -15
View File
@@ -1,17 +1,12 @@
# Production Environment Configuration
# Optimized settings for production deployment
name: "drc_names_pipeline"
version: "1.0.0"
environment: "development"
debug: true
# Processing settings
processing:
batch_size: 100_000
batch_size: 10_000
max_workers: 8
checkpoint_interval: 10
use_multiprocessing: true # Enable multiprocessing for CPU-bound tasks
use_multiprocessing: true
# Pipeline stages
stages:
@@ -20,7 +15,6 @@ stages:
#- "llm_annotation"
- "data_splitting"
# Production LLM settings
llm:
model_name: "mistral:7b"
@@ -31,14 +25,10 @@ llm:
max_concurrent_requests: 4
enable_rate_limiting: true
# Development data settings - limited dataset for faster testing
# Data handling configuration
data:
split_evaluation: true
split_by_gender: true
evaluation_fraction: 0.2
random_seed: 42
max_dataset_size: ~ # Limit to 10k records for development/testing
balance_by_sex: false # Balance male/female samples when limiting
max_dataset_size: 100_000
balance_by_sex: true
# Enhanced logging for development
logging:
config/pipeline.production.yaml
+4 -14
View File
@@ -1,17 +1,12 @@
# Production Environment Configuration
# Optimized settings for production deployment
name: "drc_names_pipeline"
version: "1.0.0"
environment: "production"
debug: false
# Production processing settings (optimized for performance)
# Processing settings
processing:
batch_size: 10_000
max_workers: 8
checkpoint_interval: 10
use_multiprocessing: true # Enable multiprocessing for CPU-bound tasks
use_multiprocessing: true
# Pipeline stages
stages:
@@ -20,7 +15,6 @@ stages:
- "llm_annotation"
- "data_splitting"
# Production LLM settings
llm:
model_name: "mistral:7b"
@@ -31,19 +25,15 @@ llm:
max_concurrent_requests: 4
enable_rate_limiting: true
# Production data settings
# Data handling configuration
data:
split_evaluation: true
split_by_gender: true
evaluation_fraction: 0.2
random_seed: 42
max_dataset_size: null
balance_by_sex: false
# Production logging (less verbose)
logging:
level: "INFO"
console_logging: false # Disable console in production
console_logging: false
file_logging: true
log_file: "pipeline.production.log"
max_log_size: 52428800 # 50MB
config/pipeline.yaml
+47 -47
View File
@@ -1,72 +1,72 @@
# DRC Names Processing Pipeline Configuration
# Main configuration file with default settings
name: "drc_names_pipeline"
version: "1.0.0"
description: "DRC Names NLP Processing Pipeline"
environment: "development"
debug: false
name: "drc_ners_pipeline" # Name of the pipeline
version: "1.0.0" # Version of the pipeline
description: "DRC NERS NLP Processing" # Description of the pipeline
environment: "development" # Environment type (development, production, etc.)
debug: false # Enable debug mode for detailed logging and error reporting
# Project directory structure
paths:
root_dir: "."
configs_dir: "./config"
data_dir: "./data/dataset"
models_dir: "./data/models"
outputs_dir: "./data/outputs"
logs_dir: "./data/logs"
checkpoints_dir: "./data/checkpoints"
root_dir: "." # Root directory of the project
configs_dir: "./config" # Directory for configuration files
data_dir: "./data/dataset" # Directory for dataset files
models_dir: "./data/models" # Directory for model files
outputs_dir: "./data/outputs" # Directory for output files
logs_dir: "./data/logs" # Directory for log files
checkpoints_dir: "./data/checkpoints" # Directory for model checkpoints
# Pipeline stages
stages:
- "data_cleaning"
- "feature_extraction"
- "llm_annotation"
- "data_splitting"
stages: # List of stages in the processing pipeline
- "data_cleaning" # Data cleaning stage
- "feature_extraction" # Feature extraction stage
- "llm_annotation" # LLM annotation stage (computational intensive)
- "data_splitting" # Data splitting stage
# Data processing configuration
processing:
batch_size: 1_000
max_workers: 4
checkpoint_interval: 5
use_multiprocessing: false
encoding_options:
batch_size: 1_000 # Size of data batches to process at once
max_workers: 4 # Number of worker threads for parallel processing
checkpoint_interval: 5 # Interval for saving checkpoints during processing
use_multiprocessing: false # Enable multiprocessing for CPU-bound tasks
encoding_options: # List of encodings to try when reading files
- "utf-8"
- "utf-16"
- "latin1"
chunk_size: 100_000
chunk_size: 100_000 # Size of data chunks to process in parallel
# LLM annotation settings
llm:
model_name: "mistral:7b"
requests_per_minute: 60
requests_per_second: 2
retry_attempts: 3
timeout_seconds: 600
max_concurrent_requests: 2
enable_rate_limiting: true
model_name: "mistral:7b" # Name of the LLM model to use
requests_per_minute: 60 # Requests per minute to the LLM service
requests_per_second: 2 # Requests per second to the LLM service
retry_attempts: 3 # Number of retry attempts for LLM requests
timeout_seconds: 600 # Timeout for LLM requests
max_concurrent_requests: 2 # Maximum concurrent requests to the LLM service
enable_rate_limiting: true # Enable rate limiting to avoid overloading the LLM service
# Data handling configuration
data:
input_file: "names.csv"
input_file: "names.csv" # Input file containing names data
output_files:
featured: "names_featured.csv"
evaluation: "names_evaluation.csv"
males: "names_males.csv"
females: "names_females.csv"
split_evaluation: true
split_by_gender: true
evaluation_fraction: 0.2
random_seed: 42
max_dataset_size: null
balance_by_sex: false
featured: "names_featured.csv" # Output file for featured data
evaluation: "names_evaluation.csv" # Output file for evaluation set
males: "names_males.csv" # Output files for male names
females: "names_females.csv" # Output files for female names
split_evaluation: true # Should the dataset be split into training and evaluation sets ?
split_by_gender: true # Should the dataset be split by gender ?
evaluation_fraction: 0.2 # Fraction of data to use for evaluation
random_seed: 42 # Random seed for reproducibility
max_dataset_size: null # Maximum size of the dataset to process, set to null for no
balance_by_sex: false # Should the dataset be balanced by sex when limiting the dataset size?
# Logging configuration
logging:
level: "INFO"
level: "INFO" # Logging level (DEBUG, INFO, WARNING, ERROR, CRITICAL)
format: "%(asctime)s - %(name)s - %(levelname)s - %(message)s"
file_logging: true
console_logging: true
log_file: "pipeline.log"
max_log_size: 10485760 # 10MB
backup_count: 5
file_logging: true # Enable logging to file
console_logging: true # Enable logging to console
log_file: "pipeline.log" # Log file name
max_log_size: 10485760 # Maximum size of log file before rotation (10MB)
backup_count: 5 # Number of backup log files to keep
config/research_templates.yaml
+129 -109
View File
@@ -1,128 +1,148 @@
# Research Experiment Configuration Templates
# These configurations can be used as starting points for different types of experiments
# Baseline Experiments Configuration
baseline_experiments:
- name: "baseline_logistic_regression_fullname"
- name: "bigru"
description: "Baseline BiGRU with full name features"
model_type: "bigru"
features: [ "full_name" ]
model_params:
max_len: 20
embedding_dim: 64
gru_units: 32
epochs: 10
batch_size: 32
tags: [ "baseline", "neural", "bigru" ]
- name: "cnn"
description: "Baseline CNN with character patterns"
model_type: "cnn"
features: [ "full_name" ]
model_params:
max_len: 20
embedding_dim: 64
filters: 64
kernel_size: 3
dropout: 0.5
epochs: 10
batch_size: 32
tags: [ "baseline", "neural", "cnn" ]
- name: "ensemble"
description: "Baseline Ensemble with multiple models"
model_type: "ensemble"
features: [ "full_name", "name_length", "word_count" ]
model_params:
base_models: [ "logistic_regression", "random_forest", "xgboost" ]
voting: "soft"
cv_folds: 5
tags: [ "baseline", "ensemble" ]
- name: "lightgbm"
description: "Baseline LightGBM with engineered features"
model_type: "lightgbm"
features: [ "full_name", "name_length", "word_count" ]
model_params:
n_estimators: 100
max_depth: -1
learning_rate: 0.1
num_leaves: 31
subsample: 0.8
colsample_bytree: 0.8
tags: [ "baseline", "lightgbm" ]
- name: "logistic_regression_fullname"
description: "Baseline logistic regression with full name"
model_type: "logistic_regression"
features: ["full_name"]
features: [ "full_name" ]
model_params:
ngram_range: [2, 5]
max_features: 10000
max_iter: 1000
tags: ["baseline", "fullname"]
tags: [ "baseline", "logistic_regression", "fullname" ]
- name: "baseline_logistic_regression_native"
- name: "logistic_regression_native"
description: "Logistic regression with native name only"
model_type: "logistic_regression"
features: ["native_name"]
features: [ "native_name" ]
model_params:
ngram_range: [2, 4]
max_features: 5000
tags: ["baseline", "native"]
tags: [ "baseline", "logistic_regression", "native" ]
- name: "baseline_rf_engineered"
description: "Random Forest with engineered features"
- name: "logistic_regression_surname"
description: "Logistic regression with surname name only"
model_type: "logistic_regression"
features: [ "surname" ]
model_params:
max_features: 5000
tags: [ "baseline", "logistic_regression", "surname" ]
- name: "lstm"
description: "Baseline LSTM with full name features"
model_type: "lstm"
features: [ "full_name" ]
model_params:
embedding_dim: 128
lstm_units: 64
epochs: 10
batch_size: 64
tags: [ "baseline", "neural", "lstm" ]
- name: "naive_bayes"
description: "Baseline Naive Bayes with full name features"
model_type: "naive_bayes"
features: [ "full_name" ]
model_params:
max_features: 5000
tags: [ "baseline", "naive_bayes" ]
- name: "random_forest"
description: "Baseline Random Forest with engineered features"
model_type: "random_forest"
features: ["name_length", "word_count", "province"]
features: [ "name_length", "word_count", "province" ]
model_params:
n_estimators: 100
max_depth: 10
tags: ["baseline", "engineered"]
min_samples_split: 2
min_samples_leaf: 1
tags: [ "baseline", "random_forest", "engineered" ]
- name: "svm"
description: "Baseline SVM with full name features"
model_type: "svm"
features: [ "full_name" ]
model_params:
C: 1.0
kernel: "rbf"
ngram_range: [ 2, 4 ]
max_features: 5000
tags: [ "baseline", "svm" ]
- name: "transformer"
description: "Baseline Transformer with attention mechanism"
model_type: "transformer"
features: [ "full_name" ]
model_params:
embedding_dim: 128
num_heads: 4
num_layers: 2
epochs: 10
batch_size: 64
tags: [ "baseline", "neural", "transformer" ]
- name: "xgboost"
description: "Baseline XGBoost with engineered features"
model_type: "xgboost"
features: [ "full_name", "name_length", "word_count" ]
model_params:
n_estimators: 100
max_depth: 6
learning_rate: 0.1
subsample: 0.8
colsample_bytree: 0.8
tags: [ "baseline", "xgboost" ]
# Advanced Experiments Configuration
advanced_experiments:
# Feature Study Configurations
feature_studies:
- name: "native_vs_surname"
description: "Compare native name vs surname effectiveness"
experiments:
- model_type: "logistic_regression"
features: ["native_name"]
tags: ["feature_study", "native"]
- model_type: "logistic_regression"
features: ["surname"]
tags: ["feature_study", "surname"]
- name: "name_parts_analysis"
description: "Analyze effectiveness of different name parts"
experiments:
- features: ["first_word"]
tags: ["name_parts", "first"]
- features: ["last_word"]
tags: ["name_parts", "last"]
- features: ["name_beginnings"]
feature_params:
beginning_length: 3
tags: ["name_parts", "beginnings"]
- features: ["name_endings"]
feature_params:
ending_length: 3
tags: ["name_parts", "endings"]
# Province-Specific Studies
province_studies:
- name: "kinshasa_study"
description: "Gender prediction for Kinshasa province"
model_type: "logistic_regression"
features: ["full_name"]
train_data_filter:
province: "kinshasa"
tags: ["province_study", "kinshasa"]
- name: "cross_province_generalization"
description: "Train on one province, test on another"
experiments:
- train_filter: {"province": "kinshasa"}
test_filter: {"province": "bas-congo"}
tags: ["generalization", "kinshasa_to_bas-congo"]
# Model Comparison Studies
model_comparisons:
- name: "model_comparison_fullname"
description: "Compare different models with full name"
base_config:
features: ["full_name"]
tags: ["model_comparison"]
models:
- model_type: "logistic_regression"
model_params:
ngram_range: [2, 5]
- model_type: "random_forest"
# Note: RF will need different feature preparation
features: ["name_length", "word_count", "province"]
# Advanced Feature Combinations
advanced_features:
- name: "multi_feature_combination"
description: "Test various feature combinations"
experiments:
- features: ["full_name", "name_length"]
tags: ["combination", "name_plus_length"]
- features: ["native_name", "surname", "province"]
tags: ["combination", "semantic_features"]
- features: ["name_beginnings", "name_endings", "word_count"]
tags: ["combination", "structural_features"]
# Hyperparameter Studies
hyperparameter_studies:
- name: "ngram_range_study"
description: "Study effect of different n-gram ranges"
base_config:
model_type: "logistic_regression"
features: ["full_name"]
tags: ["hyperparameter", "ngram"]
variants:
- model_params: {"ngram_range": [1, 3]}
- model_params: {"ngram_range": [2, 4]}
- model_params: {"ngram_range": [2, 5]}
- model_params: {"ngram_range": [3, 6]}
# Data Size Studies
data_studies:
- name: "learning_curve_study"
description: "Study performance vs training data size"
base_config:
model_type: "logistic_regression"
features: ["full_name"]
tags: ["learning_curve"]
data_sizes: [0.1, 0.25, 0.5, 0.75, 1.0] # Fractions of training data to use
# Hyperparameter Tuning Configurations
hyperparameter_tuning: