Upload folder using huggingface_hub

README.md

@@ -0,0 +1,22 @@
+---
+base_model: deepseek-ai/DeepSeek-R1-Distill-Qwen-1.5B
+library_name: peft
+pipeline_tag: text-generation
+tags:
+- latent-reasoning
+- colar
+- research
+---
+# colar-r1-cot
+
+R1 CoT 基座阶段(纯 CoT 预热, 10ep)
+
+- **Base model:** `deepseek-ai/DeepSeek-R1-Distill-Qwen-1.5B`
+- **Files:** `colar_r1_cot.ckpt`, `hparams.yaml`
+
+## Loading (PyTorch-Lightning checkpoint — NOT AutoModel-loadable)
+Weights live under the top-level key `['state_dict']` and only fit the custom CoLaR scaffold (base LLM + `[PAD]` resize + r128 q/v LoRA + a `LatentPolicy` MLP), loaded `strict=False`. Load the base separately and splice this `state_dict` in. Runtime env:
+```
+COLAR_BASE=<base>  COLAR_CKPT=colar-gsm/colar_best.ckpt  COLAR_EMB_STD=0.018  COLAR_COMPRESS=5  COLAR_MAXLAT=64  TORCH_FORCE_NO_WEIGHTS_ONLY_LOAD=1
+```
+`TORCH_FORCE_NO_WEIGHTS_ONLY_LOAD=1` is required for these older Lightning ckpts.

colar_r1_cot.ckpt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b123152c6a1b60e4bcf688755ebe548ca02310e39cf81a4f9e96f90d640f426a
+size 69792287

hparams.yaml

@@ -0,0 +1,108 @@
+model_kwargs:
+  model_id: DeepSeek-R1-Distill-Qwen-1.5B
+  sft_method: cot
+  chat_template: false
+  do_lora: true
+  lora_config:
+    r: 128
+    lora_alpha: 32
+  answer_generation_config:
+    max_new_tokens: 256
+    do_sample: true
+    top_p: 0.9
+    temperature: 1.0
+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: 1
+    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/cot
+      name: qsa-gsm_math
+      version: 20260725-205605_408408_cot_r1
+    max_epochs: 10
+  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.cot.LitCot
+    model_kwargs:
+      model_id: DeepSeek-R1-Distill-Qwen-1.5B
+      sft_method: cot
+      chat_template: false
+      do_lora: true
+      lora_config:
+        r: 128
+        lora_alpha: 32
+      answer_generation_config:
+        max_new_tokens: 256
+        do_sample: true
+        top_p: 0.9
+        temperature: 1.0
+    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: gsm_math
+    tiny_dataset: false
+    epoch_scaling: 1
+  args:
+    model: cot
+    dataset: qsa
+    trainer: default
+    devices: '0'
+    no_log: false
+    log_suffix: cot_r1
+    resume_ckpt_path: null
+    load_ckpt_path: null
+    workspace_path: /content/ws
+    do_test: false
+    test_ckpt_path: ''
+    test_times: 5
+    seed: 0
+  unkown_args:
+    dataset_name: gsm_math
+    model_id: DeepSeek-R1-Distill-Qwen-1.5B
+    batch_size: '4'
+    accumulate_grad_batches: '4'
+    max_epochs: '10'