Upload 8 files

logs/colar/qsa-gsm/colar-final/checkpoints/colar_best.ckpt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d770babf58f8b265cfc5b658862e1d3c67159dc1bdb3be5b3b7982d9514b5fa8
+size 121711010

logs/colar/qsa-gsm/colar-final/events.out.tfevents.60691.0

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9b81c001ee7cf832b7ae0d87d7996efc0a5686a13cded40084c0ff0b4eb5c1ad
+size 1298289

logs/colar/qsa-gsm/colar-final/events.out.tfevents.60691.1

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9a57b6327d0a80040cd01677060eb0266a04e9bfb71671abd726beaeb68ef3f2
+size 17551

logs/colar/qsa-gsm/colar-final/hparams.yaml

@@ -0,0 +1,103 @@
+all_config:
+  trainer:
+    target: lightning.pytorch.trainer.Trainer
+    devices:
+    - 0
+    - 1
+    - 2
+    - 3
+    - 4
+    - 5
+    - 6
+    - 7
+    max_steps: -1
+    check_val_every_n_epoch: 1
+    log_every_n_steps: 10
+    num_sanity_val_steps: 2
+    gradient_clip_val: null
+    reload_dataloaders_every_n_epochs: 0
+    accumulate_grad_batches: 1
+    precision: bf16-mixed
+    use_distributed_sampler: true
+    strategy: auto
+    logger:
+      target: lightning.pytorch.loggers.TensorBoardLogger
+      save_dir: logs/colar
+      name: qsa-gsm
+      version: colar-final
+    max_epochs: 50
+  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
+      depth: 1
+      sft_method: CoLaR
+      set_pad_as_last_token: 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: -1e-6
+        pred_embed_forward_weight: 0
+        max_compression_factor: 5
+        pred_compressed_cot: true
+        replace_r_with_auto_prob: 0
+        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:
+        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: 32
+    val_batch_size: 32
+    num_workers: 32
+    pin_memory: true
+    persistent_workers: true
+  data_module:
+    target: src.datasets.qsa.QSADataModule
+    dataset_name: gsm
+    tiny_dataset: false
+    epoch_scaling: 1

logs/colar/qsa-math/colar-rl/checkpoints/colar_best.ckpt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0af9a1be218b34c41d2de6410aa70c3e24637ea9bdc60fa9c7f3ce9774a0888e
+size 124346104

logs/colar/qsa-math/colar-rl/events.out.tfevents.14242.0

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4299501ca8fe2554dbab76ae590c22491a35d2ce5433b4f242bd39e304664d3a
+size 2803637

logs/colar/qsa-math/colar-rl/events.out.tfevents.14242.1

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b04b26dd4a3f82b23c8f733ee4aef31c2811e8d0ea207b76b75ec707d3f77d63
+size 17954

logs/colar/qsa-math/colar-rl/hparams.yaml

@@ -0,0 +1,96 @@
+all_config:
+  trainer:
+    target: lightning.pytorch.trainer.Trainer
+    devices:
+    - 0
+    max_steps: -1
+    check_val_every_n_epoch: 1
+    log_every_n_steps: 1
+    num_sanity_val_steps: 2
+    gradient_clip_val: null
+    reload_dataloaders_every_n_epochs: 0
+    accumulate_grad_batches: 1
+    precision: 32-true
+    use_distributed_sampler: true
+    strategy: auto
+    logger:
+      target: lightning.pytorch.loggers.TensorBoardLogger
+      save_dir: logs/colar
+      name: qsa-math
+      version: 
+    max_epochs: 50
+  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.latent_colar.LitCoLaR
+    model_kwargs:
+      model_id: DeepSeek-R1-Distill-Qwen-1.5B
+      depth: 1
+      sft_method: colar
+      set_pad_as_last_token: false
+      do_lora: true
+      lora_config:
+        r: 128
+        lora_alpha: 32
+      latent_cot_config:
+        ce_weight: 1
+        embed_modeling_weight: 1
+        embed_modeling_loss: nll
+        entropy_weight: 0
+        pred_embed_forward_weight: 0
+        max_compression_factor: 4
+        pred_compressed_cot: true
+        replace_r_with_auto_prob: 0.0
+        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: 2
+      answer_generation_config:
+        max_new_tokens: 16
+        do_sample: true
+        top_p: 0.9
+        temperature: 1.0
+      do_rl: true
+      rl_config:
+        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: 1.0e-06
+        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: 32
+    pin_memory: true
+    persistent_workers: true
+  data_module:
+    target: src.datasets.qsa.QSADataModule
+    dataset_name: math
+    tiny_dataset: false
+    epoch_scaling: 1