Upload training_config.yaml with huggingface_hub

training_config.yaml

@@ -0,0 +1,38 @@
+add_think_tokens: true
+chat_version: Qwen/Qwen2.5-1.5B-Instruct
+custom_name: lightr1_3_EN_1024
+dataset: Light-R1
+hf_username: mveroe
+max_length: 1024
+mixup:
+  FPR: 0.0
+  TPR: 1.0
+  generator_accuracy: 1.0
+model_name: Qwen/Qwen2.5-1.5B
+neptune_project: ethsri/label-noise
+pretrained_model: true
+seed: 42
+set_seed: false
+shuffle: true
+training_args:
+  bf16: true
+  ddp_find_unused_parameters: false
+  gradient_accumulation_steps: 4
+  hub_strategy: all_checkpoints
+  learning_rate: 5.0e-05
+  logging_first_step: true
+  logging_steps: 1
+  logging_strategy: steps
+  lr_scheduler_type: cosine
+  num_train_epochs: 3
+  optim: adafactor
+  overwrite_output_dir: true
+  per_device_train_batch_size: 4
+  push_to_hub: true
+  report_to: none
+  save_strategy: epoch
+  seed: 42
+  warmup_ratio: 0.1
+training_type: !!python/object/apply:src.configs.TrainingTypes
+- sft
+use_neptune: true