wip

.python-version

@@ -0,0 +1 @@
+3.10

.vscode/settings.json

@@ -0,0 +1,22 @@
+{
+    "workbench.colorCustomizations": {
+        "activityBar.activeBackground": "#ebae44",
+        "activityBar.background": "#ebae44",
+        "activityBar.foreground": "#15202b",
+        "activityBar.inactiveForeground": "#15202b99",
+        "activityBarBadge.background": "#0f9061",
+        "activityBarBadge.foreground": "#e7e7e7",
+        "commandCenter.border": "#15202b99",
+        "sash.hoverBorder": "#ebae44",
+        "statusBar.background": "#e49a18",
+        "statusBar.foreground": "#15202b",
+        "statusBarItem.hoverBackground": "#b67b13",
+        "statusBarItem.remoteBackground": "#e49a18",
+        "statusBarItem.remoteForeground": "#15202b",
+        "titleBar.activeBackground": "#e49a18",
+        "titleBar.activeForeground": "#15202b",
+        "titleBar.inactiveBackground": "#e49a1899",
+        "titleBar.inactiveForeground": "#15202b99"
+    },
+    "peacock.remoteColor": "#e49a18"
+}

README.md

@@ -1,3 +1,14 @@
+I'm using this to train some simple base -> SFT models for my  work
+
+
+```sh
+uv sync --no-build-isolation-package flash-attn
+MAX_JOBS=10 pip install flash-attn --no-build-isolation
+```
+
+
+Old readme
+----
 <p align="center">
   <img src="https://raw.githubusercontent.com/huggingface/alignment-handbook/main/assets/handbook.png">
 </p>

pyproject.toml

@@ -0,0 +1,56 @@
+[project]
+name = "alignment-handbook"
+version = "0.1.0"
+description = "Add your description here"
+readme = "README.md"
+requires-python = ">=3.10"
+
+dependencies = [
+    "accelerate>=0.29.2",
+    "bitsandbytes>=0.43.0",
+    "black>=24.4.2",
+    "datasets>=2.18.0",
+    "deepspeed>=0.14.4",
+    "einops>=0.6.1",
+    "evaluate==0.4.0",
+    "flake8>=6.0.0",
+    "hf-doc-builder>=0.4.0",
+    "hf_transfer>=0.1.4",
+    "huggingface-hub>=0.19.2,<1.0",
+    "isort>=5.12.0",
+    "ninja>=1.11.1",
+    "numpy>=1.24.2",
+    "packaging>=23.0",
+    "parameterized>=0.9.0",
+    "peft>=0.9.0",
+    "protobuf<=3.20.2", # Needed to avoid conflicts with `transformers`
+    "pytest",
+    "safetensors>=0.3.3",
+    "sentencepiece>=0.1.99",
+    "scipy",
+    "tensorboard",
+    "torch>=2.1.2",
+    "transformers>=4.39.3",
+    "trl>=0.9.6,<0.13.0",
+    "jinja2>=3.0.0",
+    "tqdm>=4.64.1",
+    "wheel>=0.45.1",
+    "setuptools>=80.9.0",
+    "hatchling>=1.27.0",
+    "editables>=0.5",
+]
+
+
+[dependency-groups]
+dev = [
+    "pytest",
+    "ipykernel>=6.29.5",
+]
+
+[build-system]
+requires = ["hatchling"]
+build-backend = "hatchling.build"
+
+[tool.hatch.build.targets.wheel]
+packages = ["src"]
+

recipes/fromSimPO/llama-3-2-1b-base-sft copy.yaml

@@ -0,0 +1,48 @@
+# Model arguments
+model_name_or_path: NousResearch/Llama-3.2-1B
+model_revision: main
+torch_dtype: bfloat16
+attn_implementation: flash_attention_2
+
+# Data training arguments
+chat_template: "{% set loop_messages = messages %}{% for message in loop_messages %}{% set content = '<|start_header_id|>' + message['role'] + '<|end_header_id|>\n\n'+ message['content'] | trim + '<|eot_id|>' %}{% if loop.index0 == 0 %}{% set content = bos_token + content %}{% endif %}{{ content }}{% endfor %}{% if add_generation_prompt %}{{ '<|start_header_id|>assistant<|end_header_id|>\n\n' }}{% endif %}"
+dataset_mixer:
+  wassname/ultrachat_200k_filtered: 1.0
+dataset_splits:
+- train_sft
+- test_sft
+preprocessing_num_workers: 12
+
+# SFT trainer config
+bf16: true
+do_eval: true
+evaluation_strategy: steps
+eval_steps: 200
+gradient_accumulation_steps: 32
+gradient_checkpointing: true
+gradient_checkpointing_kwargs:
+  use_reentrant: False
+hub_model_id: llama-3.2-1b-sft
+hub_strategy: every_save
+learning_rate: 2.0e-04
+log_level: info
+logging_steps: 5  
+logging_strategy: steps
+lr_scheduler_type: cosine
+max_seq_length: 2048
+max_steps: -1
+num_train_epochs: 3
+output_dir: /workspace/checkpoints_new/llama-3-2-1b-sft
+run_name: llama-3-2-1b-sft
+overwrite_output_dir: true
+per_device_eval_batch_size: 8
+per_device_train_batch_size: 8
+push_to_hub: false
+remove_unused_columns: true
+report_to:
+- wandb
+save_strategy: "steps"
+save_steps: 1000000
+save_total_limit: 1
+seed: 42
+warmup_ratio: 0.1

recipes/fromSimPO/llama-3-2-1b-base-sft.yaml

@@ -0,0 +1,48 @@
+# Model arguments
+model_name_or_path: NousResearch/Llama-3.2-1B
+model_revision: main
+torch_dtype: bfloat16
+attn_implementation: flash_attention_2
+
+# Data training arguments
+chat_template: "{% set loop_messages = messages %}{% for message in loop_messages %}{% set content = '<|start_header_id|>' + message['role'] + '<|end_header_id|>\n\n'+ message['content'] | trim + '<|eot_id|>' %}{% if loop.index0 == 0 %}{% set content = bos_token + content %}{% endif %}{{ content }}{% endfor %}{% if add_generation_prompt %}{{ '<|start_header_id|>assistant<|end_header_id|>\n\n' }}{% endif %}"
+dataset_mixer:
+  wassname/ultrachat_200k_filtered: 1.0
+dataset_splits:
+- train_sft
+- test_sft
+preprocessing_num_workers: 12
+
+# SFT trainer config
+bf16: true
+do_eval: true
+evaluation_strategy: steps
+eval_steps: 200
+gradient_accumulation_steps: 32
+gradient_checkpointing: true
+gradient_checkpointing_kwargs:
+  use_reentrant: False
+hub_model_id: llama-3.2-1b-sft
+hub_strategy: every_save
+learning_rate: 2.0e-04
+log_level: info
+logging_steps: 5  
+logging_strategy: steps
+lr_scheduler_type: cosine
+max_seq_length: 2048
+max_steps: -1
+num_train_epochs: 3
+output_dir: /workspace/checkpoints_new/llama-3-2-1b-sft
+run_name: llama-3-2-1b-sft
+overwrite_output_dir: true
+per_device_eval_batch_size: 8
+per_device_train_batch_size: 8
+push_to_hub: false
+remove_unused_columns: true
+report_to:
+- wandb
+save_strategy: "steps"
+save_steps: 1000000
+save_total_limit: 1
+seed: 42
+warmup_ratio: 0.1

recipes/fromSimPO/llama-3-2-3b-base-sft.yaml

@@ -0,0 +1,48 @@
+# Model arguments
+model_name_or_path: tanliboy/Llama-3.2-3B
+model_revision: main
+torch_dtype: bfloat16
+attn_implementation: flash_attention_2
+
+# Data training arguments
+chat_template: "{% set loop_messages = messages %}{% for message in loop_messages %}{% set content = '<|start_header_id|>' + message['role'] + '<|end_header_id|>\n\n'+ message['content'] | trim + '<|eot_id|>' %}{% if loop.index0 == 0 %}{% set content = bos_token + content %}{% endif %}{{ content }}{% endfor %}{% if add_generation_prompt %}{{ '<|start_header_id|>assistant<|end_header_id|>\n\n' }}{% endif %}"
+dataset_mixer:
+  wassname/ultrachat_200k_filtered: 1.0
+dataset_splits:
+- train_sft
+- test_sft
+preprocessing_num_workers: 12
+
+# SFT trainer config
+bf16: true
+do_eval: true
+evaluation_strategy: steps
+eval_steps: 200
+gradient_accumulation_steps: 8
+gradient_checkpointing: true
+gradient_checkpointing_kwargs:
+  use_reentrant: False
+hub_model_id: llama-3.2-3b-sft
+hub_strategy: every_save
+learning_rate: 2.0e-05
+log_level: info
+logging_steps: 5  
+logging_strategy: steps
+lr_scheduler_type: cosine
+max_seq_length: 2048
+max_steps: -1
+num_train_epochs: 1
+output_dir: /workspace/checkpoints_new/llama-3-2-3b-sft
+run_name: llama-3-2-3b-sft
+overwrite_output_dir: true
+per_device_eval_batch_size: 3
+per_device_train_batch_size: 2
+push_to_hub: false
+remove_unused_columns: true
+report_to:
+- wandb
+save_strategy: "steps"
+save_steps: 1000000
+save_total_limit: 1
+seed: 42
+warmup_ratio: 0.1

recipes/fromSimPO/llama-3-8b-base-sft.yaml

@@ -0,0 +1,48 @@
+# Model arguments
+model_name_or_path: NousResearch/Meta-Llama-3-8B
+model_revision: main
+torch_dtype: bfloat16
+attn_implementation: flash_attention_2
+
+# Data training arguments
+chat_template: "{% set loop_messages = messages %}{% for message in loop_messages %}{% set content = '<|start_header_id|>' + message['role'] + '<|end_header_id|>\n\n'+ message['content'] | trim + '<|eot_id|>' %}{% if loop.index0 == 0 %}{% set content = bos_token + content %}{% endif %}{{ content }}{% endfor %}{% if add_generation_prompt %}{{ '<|start_header_id|>assistant<|end_header_id|>\n\n' }}{% endif %}"
+dataset_mixer:
+  wassname/ultrachat_200k_filtered: 1.0
+dataset_splits:
+- train_sft
+- test_sft
+preprocessing_num_workers: 12
+
+# SFT trainer config
+bf16: true
+do_eval: true
+evaluation_strategy: steps
+eval_steps: 200
+gradient_accumulation_steps: 4
+gradient_checkpointing: true
+gradient_checkpointing_kwargs:
+  use_reentrant: False
+hub_model_id: zephyr-7b-sft-full
+hub_strategy: every_save
+learning_rate: 2.0e-05
+log_level: info
+logging_steps: 5  
+logging_strategy: steps
+lr_scheduler_type: cosine
+max_seq_length: 2048
+max_steps: -1
+num_train_epochs: 1
+output_dir: /scratch/gpfs/DANQIC/ym0081/checkpoints_new/llama-3-8b-sft
+run_name: llama-3-8b-sft
+overwrite_output_dir: true
+per_device_eval_batch_size: 8
+per_device_train_batch_size: 8
+push_to_hub: false
+remove_unused_columns: true
+report_to:
+- wandb
+save_strategy: "steps"
+save_steps: 1000000
+save_total_limit: 1
+seed: 42
+warmup_ratio: 0.1