# CNN/DailyMail Pointwise Sentence Scorer Run

- model_size: `base`
- model_name: `BAAI/bge-m3`
- resolved_model_name: `/lustre/fsmisc/dataset/HuggingFace_Models/BAAI/bge-m3`
- target_mode: `model_balanced`
- target_field: `salience_score_model_balanced`
- loss_type: `mse`
- score_transform: `identity`
- article_balanced: `False`
- multihead: `False`
- multihead_names: `[]`
- score_aggregation: `scalar`
- eval_loss_semantics: `mse`
- input_field: `model_input_text`
- train_file: `data/CNN_DM/representative_subset/salience_generation/recovery_f2_all6/sentence_scorer_data/modernbert_pointwise/train.jsonl`
- dev_file: `data/CNN_DM/representative_subset/salience_generation/recovery_f2_all6/sentence_scorer_data/modernbert_pointwise/dev.jsonl`
- train_rows: `59643`
- dev_rows: `6793`
- train_articles: `1000`
- dev_articles: `200`
- max_length: `4608`
- attn_implementation: `eager`
- learning_rate: `1e-05`
- weight_decay: `0.01`
- num_train_epochs: `3.0`
- warmup_ratio: `0.06`
- per_device_train_batch_size: `1`
- per_device_eval_batch_size: `2`
- gradient_accumulation_steps: `16`
- effective_train_batch_size: `16`
- eval_strategy: `steps`
- save_strategy: `steps`
- early_stopping_patience: `5`
- bf16: `True`
- fp16: `False`
- gradient_checkpointing: `True`

## Eval Metrics

- epoch: `2.3068926781013697`
- eval_article_ndcg_at_3: `0.8912648201819963`
- eval_article_ndcg_at_5: `0.8810350116689233`
- eval_article_spearman: `0.6913812102325307`
- eval_loss: `0.03421084210276604`
- eval_mae: `0.13758335639960514`
- eval_mse: `0.03421084379263475`
- eval_pearson: `0.7992694624484324`
- eval_rmse: `0.18496173602298058`
- eval_runtime: `136.7425`
- eval_samples_per_second: `49.677`
- eval_spearman: `0.7975893083370461`
- eval_steps_per_second: `24.842`