Start agent traces

README_AGENTS.md

@@ -0,0 +1,21 @@
+Launch:
+```bash
+sbatch --nodes=1 slurm/train.slurm --model SmolLM2-1.7B-Instruct --task sft --config agent --accelerator zero3
+```
+Refers to the config  recipes/SmolLM2-1.7B-Instruct/sft/config_agent.yaml
+zero3 is one of the accelerate configs in recipes/accelerate_configs
+
+
+### VLM training
+
+Launch in multi GPU:
+```bash
+sbatch --qos=high --nodes=1 slurm/train.slurm --model Qwen2.5-VL-3B-Instruct --task sft --config agent --accelerator zero3
+```
+
+🛑 For me the above fails because of NCCL issues, I launch it in single-GPU mode as follows:
+```bash
+sbatch slurm/trainsingle.slurm --model Qwen2.5-VL-3B-Instruct --task sft --config agent
+```
+
+The config is located under recipes/Qwen2.5-VL-3B-Instruct/sft/config_agent.yaml

recipes/Qwen2.5-3B-Instruct/sft/config.yaml

@@ -0,0 +1,46 @@
+# Model arguments
+# You can download the model and manually change the rope to 300k/500k and max_position_embeddings to 32768
+model_name_or_path: HuggingFaceTB/SmolLM2-1.7B-Instruct
+model_revision: main
+torch_dtype: bfloat16
+attn_implementation: sdpa
+
+# Data training arguments
+dataset_name: open-r1/OpenR1-Math-220k
+dataset_num_proc: 48
+
+#SFT hyperparam
+max_length: 8192 # You can set this to 32768 if you change the rope, but you need to change the config.json file 
+weight_decay: 0.0001
+optim: adamw_torch
+lr_scheduler_type: linear
+warmup_ratio: 0.1
+learning_rate: 5.0e-05
+gradient_accumulation_steps: 2
+per_device_eval_batch_size: 4
+per_device_train_batch_size: 4 # Change this depending on the context length of the model to keep a 500M GBS. 
+
+# SFT trainer config
+max_steps: -1
+num_train_epochs: 3
+bf16: true
+do_eval: false
+eval_strategy: 'no'
+gradient_checkpointing: true
+gradient_checkpointing_kwargs:
+  use_reentrant: false
+hub_model_id: OpenR1-Qwen-7B-SFT
+hub_strategy: every_save
+log_level: info
+logging_steps: 5
+logging_strategy: steps
+packing: true
+output_dir: data/OpenR1-Qwen-7B-SFT
+overwrite_output_dir: true
+push_to_hub: true
+report_to:
+- wandb
+save_strategy: "steps"
+save_steps: 500
+save_total_limit: 1
+seed: 42

recipes/Qwen2.5-3B-Instruct/sft/config_agent.yaml

@@ -0,0 +1,46 @@
+# Model arguments
+# You can download the model and manually change the rope to 300k/500k and max_position_embeddings to 32768
+model_name_or_path: Qwen/Qwen2.5-3B-Instruct
+model_revision: main
+torch_dtype: bfloat16
+attn_implementation: sdpa
+
+# Data training arguments
+dataset_name: smolagents/training-traces
+dataset_num_proc: 48
+
+#SFT hyperparam
+max_length: 8192 # You can set this to 32768 if you change the rope, but you need to change the config.json file 
+weight_decay: 0.0001
+optim: adamw_torch
+lr_scheduler_type: linear
+warmup_ratio: 0.1
+learning_rate: 4.0e-05
+gradient_accumulation_steps: 1
+per_device_eval_batch_size: 4
+per_device_train_batch_size: 2 # Change this depending on the context length of the model to keep a 500M GBS. 
+
+# SFT trainer config
+max_steps: -1
+num_train_epochs: 2
+bf16: true
+do_eval: false
+eval_strategy: 'no'
+gradient_checkpointing: true
+gradient_checkpointing_kwargs:
+  use_reentrant: false
+hub_model_id: oR1-Qwen-3B-Agentic-e2-lr4e-b2
+hub_strategy: every_save
+log_level: info
+logging_steps: 5
+logging_strategy: steps
+packing: true
+output_dir: data/oR1-Qwen-3B-Agentic-e2-lr4e-b2
+overwrite_output_dir: true
+push_to_hub: true
+report_to:
+- wandb
+save_strategy: "steps"
+save_steps: 500
+save_total_limit: 1
+seed: 42

recipes/Qwen2.5-VL-3B-Instruct/sft/config_agent.yaml

@@ -0,0 +1,76 @@
+# Model arguments
+# You can download the model and manually change the rope to 300k/500k and max_position_embeddings to 32768
+model_name_or_path: Qwen/Qwen2.5-VL-3B-Instruct
+vision_model: true
+model_revision: main
+torch_dtype: bfloat16
+attn_implementation: sdpa
+
+# Data training arguments
+dataset_name: smolagents/aguvis-stage-2
+dataset_num_proc: 48
+
+#SFT hyperparam
+max_length: 32768
+optim: adamw_torch
+lr_scheduler_type: cosine_with_min_lr
+lr_scheduler_kwargs:
+  min_lr_rate: 0.1
+max_grad_norm: 0.2
+warmup_ratio: 0.03
+learning_rate: 1.0e-05
+gradient_accumulation_steps: 8
+per_device_eval_batch_size: 4
+per_device_train_batch_size: 4 # Change this depending on the context length of the model to keep a 500M GBS. 
+
+single_gpu: true
+
+# SFT trainer config
+max_steps: -1
+num_train_epochs: 1
+bf16: true
+do_eval: false
+eval_strategy: 'no'
+gradient_checkpointing: true
+gradient_checkpointing_kwargs:
+  use_reentrant: false
+hub_model_id: smolagents/Qwen2.5-VL-3B-Instruct-Agentic
+hub_strategy: end
+push_to_hub: true
+log_level: info
+logging_steps: 5
+logging_strategy: steps
+output_dir: data/smolagents-Qwen2.5-VL-3B-Instruct-Agentic
+overwrite_output_dir: true
+report_to:
+- wandb
+save_strategy: "epoch"
+save_steps: 1
+save_total_limit: 1
+seed: 42
+
+dataset_mixture:
+  datasets:                     # List of datasets to include in the mixture
+    - id: smolagents/aguvis-stage-2  # Hub dataset ID
+      config: mind2web     # Name of the dataset config
+      split: train            # Split to use from the dataset
+      columns:                  # Columns to keep
+        - images
+        - texts
+      weight: 1.
+    - id: smolagents/aguvis-stage-2
+      config: guiact-web-single
+      split: train
+      columns:
+        - images
+        - texts
+      weight: 1.
+    - id: smolagents/aguvis-stage-2
+      config: guiact-web-multi
+      split: train
+      columns:
+        - images
+        - texts
+      weight: 1.
+  seed: 42                      # Seed for shuffling the combined dataset
+  test_split_size: 0.1   

recipes/SmolLM2-1.7B-Instruct/sft/config_agent.yaml

@@ -0,0 +1,45 @@
+# Model arguments
+# You can download the model and manually change the rope to 300k/500k and max_position_embeddings to 32768
+model_name_or_path: HuggingFaceTB/SmolLM2-1.7B-Instruct
+model_revision: main
+torch_dtype: bfloat16
+attn_implementation: sdpa
+
+# Data training arguments
+dataset_name: smolagents/codeagent-traces
+dataset_num_proc: 48
+
+#SFT hyperparam
+max_length: 8192 # You can set this to 32768 if you change the rope, but you need to change the config.json file 
+weight_decay: 0.0001
+optim: adamw_torch
+lr_scheduler_type: linear
+warmup_ratio: 0.1
+learning_rate: 5.0e-05
+gradient_accumulation_steps: 2
+per_device_eval_batch_size: 4
+per_device_train_batch_size: 4 # Change this depending on the context length of the model to keep a 500M GBS. 
+
+# SFT trainer config
+max_steps: -1
+num_train_epochs: 1
+bf16: true
+do_eval: false
+eval_strategy: 'no'
+gradient_checkpointing: true
+gradient_checkpointing_kwargs:
+  use_reentrant: false
+hub_model_id: OpenR1-SmolLM2-1.7B-Instruct-Agentic
+hub_strategy: every_save
+log_level: info
+logging_steps: 5
+logging_strategy: steps
+output_dir: data/OpenR1-SmolLM2-1.7B-Instruct-Agentic
+overwrite_output_dir: true
+push_to_hub: true
+report_to:
+- wandb
+save_strategy: "steps"
+save_steps: 500
+save_total_limit: 1
+seed: 42