| { |
| "model_name": "brendaogutu/mul-sw-en-translation-phase1", |
| "language_pair": "sw-en", |
| "language_name": "Swahili", |
| "dataset_config": { |
| "primary_dataset": "custom", |
| "custom_datasets": [ |
| { |
| "path": "/tmp/dataset/NLLB.en-sw-4-cleaned_clean.jsonl", |
| "name": "nllb_swahili_3", |
| "weight": 1.0, |
| "priority": "high" |
| }, |
| { |
| "path": "/tmp/dataset/NLLB.en-sw-5-cleaned_clean.jsonl", |
| "name": "nllb_swahili_4", |
| "weight": 1.0, |
| "priority": "high" |
| } |
| ], |
| "validation_split": 0.1, |
| "test_split": 0.05, |
| "max_samples": null, |
| "shuffle": true, |
| "seed": 42, |
| "max_length": 500, |
| "filter_length_ratio": true, |
| "max_length_ratio": 2.5, |
| "min_length_ratio": 0.4, |
| "tokenization": { |
| "batch_size": 1000, |
| "num_proc": 1, |
| "use_cache": true |
| } |
| }, |
| "training_config": { |
| "learning_rate": 2e-05, |
| "batch_size": 256, |
| "gradient_accumulation_steps": 1, |
| "num_epochs": 6, |
| "max_length": 256, |
| "weight_decay": 0.01, |
| "warmup_steps": 0, |
| "warmup_ratio": 0.1, |
| "lr_scheduler": "cosine", |
| "save_strategy": "steps", |
| "save_steps": 2000, |
| "eval_strategy": "steps", |
| "eval_steps": 2000, |
| "logging_steps": 50, |
| "early_stopping_patience": 5, |
| "early_stopping_threshold": 0.001, |
| "mixed_precision": "fp16", |
| "gradient_checkpointing": false, |
| "dataloader_num_workers": 4, |
| "load_best_model_at_end": true, |
| "metric_for_best_model": "eval_bleu", |
| "greater_is_better": true, |
| "label_smoothing": 0.1, |
| "max_grad_norm": 1.0, |
| "adam_beta1": 0.9, |
| "adam_beta2": 0.999, |
| "adam_epsilon": 1e-08, |
| "tensorboard_enabled": true, |
| "per_device_eval_batch_size": 64 |
| }, |
| "generation_config": { |
| "max_length": 256, |
| "min_length": 1, |
| "num_beams": 5, |
| "length_penalty": 0.6, |
| "early_stopping": true, |
| "no_repeat_ngram_size": 4, |
| "repetition_penalty": 1.5, |
| "do_sample": false |
| }, |
| "evaluation_config": { |
| "metrics": [ |
| "bleu", |
| "chrf", |
| "comet_qe" |
| ], |
| "compute_comet_during_training": false, |
| "test_size": 1000, |
| "max_eval_samples": 2000, |
| "save_predictions": true, |
| "batch_size": 32 |
| }, |
| "mlflow_config": { |
| "tracking_uri": "https://proceedings-graphic-pressed-responded.trycloudflare.com", |
| "experiment_name": "swahili-english-translation", |
| "run_name": "sw-en-phase2-nllb-3-4", |
| "log_models": true, |
| "log_artifacts": true, |
| "log_metrics": true |
| }, |
| "deployment": { |
| "model_output_dir": "models/finetuned-sw-en-phase2", |
| "save_tokenizer": true, |
| "save_config": true, |
| "create_model_card": true, |
| "push_to_hub": true, |
| "hub_model_id": "brendaogutu/mul-sw-en-translation-phase2", |
| "push_strategy": "checkpoint" |
| }, |
| "system_config": { |
| "cache_dir": ".cache", |
| "use_cuda": true, |
| "cuda_device": "cuda:0", |
| "seed": 42, |
| "deterministic": true, |
| "num_workers": 4 |
| } |
| } |
