| model_kwargs: |
| model_id: Llama-3.2-1B-Instruct |
| sft_method: colar |
| chat_template: false |
| do_lora: true |
| lora_config: |
| r: 128 |
| lora_alpha: 32 |
| latent_cot_config: |
| ce_weight: 1 |
| embed_modeling_weight: 1 |
| embed_modeling_loss: mse |
| entropy_weight: 0 |
| pred_embed_forward_weight: 0 |
| max_compression_factor: 5 |
| pred_compressed_cot: true |
| sqrt_mean: true |
| latent_policy_config: |
| lp_determinisitc: false |
| lp_intermediate_size: 2048 |
| latent_generation_config: |
| max_n_latent_forward: 64 |
| latent_temperature: 1.0 |
| compression_factor: 5 |
| answer_generation_config: |
| max_new_tokens: 16 |
| do_sample: true |
| top_p: 0.9 |
| temperature: 1.0 |
| do_rl: false |
| rl_config: |
| average_per_token_loss: false |
| random_speed_in_group: false |
| filter_dataset: false |
| exp_batch_size: 8 |
| group_size: 8 |
| punish_latent_length: false |
| clip_grad_norm: 1.0 |
| clip_eps: 0.2 |
| use_latent_loss: true |
| use_answer_loss: true |
| n_train_samples_per_epoch: 512 |
| training_kwargs: |
| optimizer: |
| target: torch.optim.AdamW |
| lr: 0.0001 |
| weight_decay: 0.01 |
| use_scheduler: false |
| scheduler: |
| target: constant_schedule_with_warmup |
| warmup_steps: 1000 |
| all_config: |
| trainer: |
| target: lightning.pytorch.trainer.Trainer |
| devices: |
| - 0 |
| max_steps: -1 |
| check_val_every_n_epoch: 5 |
| log_every_n_steps: 10 |
| num_sanity_val_steps: 2 |
| gradient_clip_val: 1.0 |
| reload_dataloaders_every_n_epochs: 0 |
| accumulate_grad_batches: 4 |
| precision: bf16-mixed |
| use_distributed_sampler: true |
| strategy: auto |
| logger: |
| target: lightning.pytorch.loggers.TensorBoardLogger |
| save_dir: logs/colar |
| name: qsa-coding_mix |
| version: 20260802-200100_983467_coding_gsmwarm |
| max_epochs: 25 |
| callbacks: |
| - target: lightning.pytorch.callbacks.ModelCheckpoint |
| save_last: true |
| save_top_k: 3 |
| mode: max |
| monitor: monitor |
| auto_insert_metric_name: false |
| filename: epoch{epoch}__step{step}__monitor{monitor:.3f} |
| save_weights_only: true |
| seed: null |
| model: |
| target: src.models.colar.LitCoLaR |
| model_kwargs: |
| model_id: Llama-3.2-1B-Instruct |
| sft_method: colar |
| chat_template: false |
| do_lora: true |
| lora_config: |
| r: 128 |
| lora_alpha: 32 |
| latent_cot_config: |
| ce_weight: 1 |
| embed_modeling_weight: 1 |
| embed_modeling_loss: mse |
| entropy_weight: 0 |
| pred_embed_forward_weight: 0 |
| max_compression_factor: 5 |
| pred_compressed_cot: true |
| sqrt_mean: true |
| latent_policy_config: |
| lp_determinisitc: false |
| lp_intermediate_size: 2048 |
| latent_generation_config: |
| max_n_latent_forward: 64 |
| latent_temperature: 1.0 |
| compression_factor: 5 |
| answer_generation_config: |
| max_new_tokens: 16 |
| do_sample: true |
| top_p: 0.9 |
| temperature: 1.0 |
| do_rl: false |
| rl_config: |
| average_per_token_loss: false |
| random_speed_in_group: false |
| filter_dataset: false |
| exp_batch_size: 8 |
| group_size: 8 |
| punish_latent_length: false |
| clip_grad_norm: 1.0 |
| clip_eps: 0.2 |
| use_latent_loss: true |
| use_answer_loss: true |
| n_train_samples_per_epoch: 512 |
| training_kwargs: |
| optimizer: |
| target: torch.optim.AdamW |
| lr: 0.0001 |
| weight_decay: 0.01 |
| use_scheduler: false |
| scheduler: |
| target: constant_schedule_with_warmup |
| warmup_steps: 1000 |
| dataloader: |
| batch_size: 4 |
| val_batch_size: 32 |
| num_workers: 8 |
| pin_memory: true |
| persistent_workers: true |
| data_module: |
| target: src.datasets.qsa.QSADataModule |
| dataset_name: coding_mix |
| tiny_dataset: false |
| epoch_scaling: 1 |
| args: |
| model: colar |
| dataset: qsa |
| trainer: default |
| devices: '0' |
| no_log: false |
| log_suffix: coding_gsmwarm |
| resume_ckpt_path: null |
| load_ckpt_path: /content/colar_hf/logs/colar/qsa-gsm/colar-final/checkpoints/colar_best.ckpt |
| workspace_path: /content/ws |
| do_test: false |
| test_ckpt_path: '' |
| test_times: 5 |
| seed: 0 |
| unkown_args: |
| dataset_name: coding_mix |
| model_id: Llama-3.2-1B-Instruct |
| batch_size: '4' |
| accumulate_grad_batches: '4' |
| max_epochs: '25' |
| check_val_every_n_epoch: '5' |
|
|
