| { |
| "experiment_name": "m2_metadata_category_classifier_draft", |
| "display_name": "Metadata Category Classifier", |
| "model_id": "M2", |
| "tier": "T1_metadata", |
| "backbone": { |
| "type": "encoder", |
| "base_model": "allenai/scibert_scivocab_uncased", |
| "adapter_type": "none", |
| "parameters_millions": 110, |
| "cache_dir": "/data/checkpoints", |
| "load_in_8bit": false, |
| "load_in_4bit": false |
| }, |
| "dataset": { |
| "sources": [ |
| "arxiv_metadata" |
| ], |
| "filters": { |
| "years": [ |
| 2000, |
| 2025 |
| ] |
| }, |
| "construction": { |
| "input_fields": [ |
| "title", |
| "abstract" |
| ], |
| "target_fields": [ |
| "categories" |
| ], |
| "max_samples": 4000, |
| "train_val_test_split": [ |
| 0.9, |
| 0.1, |
| 0.0 |
| ], |
| "shuffling_seed": 42 |
| }, |
| "tokenization": { |
| "tokenizer_name": "allenai/scibert_scivocab_uncased", |
| "max_source_tokens": 512, |
| "max_target_tokens": 128, |
| "truncate_strategy": "longest_first" |
| }, |
| "cache_dir": "/data/checkpoints", |
| "hardening": { |
| "max_per_label": 128, |
| "max_labels": 32, |
| "min_per_label": 16 |
| } |
| }, |
| "training": { |
| "batch_size": 8, |
| "gradient_accumulation_steps": 1, |
| "num_epochs": 3, |
| "learning_rate": 5e-05, |
| "optimizer": "adamw", |
| "weight_decay": 0.01, |
| "warmup_steps": 100, |
| "precision": "bf16", |
| "gradient_checkpointing": false, |
| "objective": "cross_entropy", |
| "checkpoint_dir": "models/checkpoints/M2", |
| "cache_dir": "/data/checkpoints", |
| "finetune_strategy": "full_finetune", |
| "use_hf_trainer": true, |
| "max_steps": 1000, |
| "allow_cpu_trainer": true, |
| "force_cpu": false, |
| "save_steps": 200, |
| "logging_steps": 50, |
| "eval_strategy": "no", |
| "eval_steps": 200, |
| "evaluation_strategy": "no", |
| "metric_for_best_model": "eval_loss", |
| "load_best_model_at_end": false, |
| "model_type": "classifier", |
| "resume_from_checkpoint": false, |
| "num_labels": 32 |
| }, |
| "evaluation": { |
| "metrics": [ |
| "accuracy", |
| "macro_f1" |
| ], |
| "eval_interval_steps": 200, |
| "early_stopping": { |
| "metric": "validation_loss", |
| "patience": 2 |
| } |
| }, |
| "inference": { |
| "max_new_tokens": 256, |
| "temperature": 0.7, |
| "top_p": 0.95, |
| "beam_size": 1 |
| }, |
| "compute": { |
| "gpus": [ |
| { |
| "type": "RTX_3090", |
| "count": 2, |
| "memory_gb": 24 |
| } |
| ], |
| "estimated_gpu_hours": 0, |
| "distributed_strategy": "ddp" |
| } |
| } |
|
|
