Add Smollm

README.md

@@ -19,6 +19,7 @@ However, we know from the [InstructGPT](https://huggingface.co/papers/2203.02155
 The Alignment Handbook aims to fill that gap by providing the community with a series of robust training recipes that span the whole pipeline.
 
 ## News 🗞️
+* **August 18, 2024**: We release SmolLM-Instruct v0.2, along with the [recipe](recipes/smollm/README.md)  to fine-tuning small LLMs 💻
 * **April 12, 2024**: We release Zephyr 141B (A35B), in collaboration with Argilla and Kaist AI, along with the recipe to fine-tune Mixtral 8x22B with ORPO 🪁
 * **March 12, 2024:** We release StarChat2 15B, along with the recipe to train capable coding assistants 🌟
 * **March 1, 2024:** We release Zephyr 7B Gemma, which is a new recipe to align Gemma 7B with RLAIF 🔥

recipes/smollm/README.md

@@ -0,0 +1,19 @@
+
+# Instructions to train SmolLM-Instruct
+
+We build the [SmolLM-Instruct](https://huggingface.co/collections/HuggingFaceTB/smollm-6695016cad7167254ce15966) (v0.2) models (135M, 360M and 1.7B) by doing SFT on a mix of these datasets:
+- a dataset of 2k simple everyday conversations we generated by llama3.1-70B [everyday-conversations-llama3.1-2k](https://huggingface.co/datasets/HuggingFaceTB/everyday-conversations-llama3.1-2k/)
+- [Magpie-Pro-300K-Filtered](https://huggingface.co/datasets/Magpie-Align/Magpie-Pro-300K-Filtered)
+- [StarCoder2-Self-OSS-Instruct](https://huggingface.co/datasets/bigcode/self-oss-instruct-sc2-exec-filter-50k)
+- A small subset of [OpenHermes-2.5](https://huggingface.co/datasets/teknium/OpenHermes-2.5)
+
+## Setup
+
+Follow the installation instructions in https://github.com/huggingface/alignment-handbook/tree/main?tab=readme-ov-file#installation-instructions 
+
+## Training
+We train the models on 8 GPUs using the following command:
+
+```shell
+ACCELERATE_LOG_LEVEL=info accelerate launch --config_file recipes/accelerate_configs/deepspeed_zero3.yaml scripts/run_sft.py recipes/smollm/sft/config.yaml
+```

recipes/smollm/sft/config.yaml

@@ -0,0 +1,53 @@
+# Model arguments
+model_name_or_path: HuggingFaceTB/SmolLM-360M
+model_revision: main
+tokenizer_name_or_path: HuggingFaceTB/SmolLM-360M-Instruct # Custom tokenizer with <|im_start|> and <|im_end|> tokens
+torch_dtype: bfloat16
+use_flash_attention_2: true
+
+# Data training arguments
+dataset_mixer:
+  HuggingFaceTB/Magpie-Pro-300K-Filtered-H4: 1.0 
+  HuggingFaceTB/self-oss-instruct-sc2-H4: 1.0 
+  HuggingFaceTB/OpenHermes-2.5-H4: 0.001 
+  HuggingFaceTB/everyday-conversations-llama3.1-2k: 1.0 
+  HuggingFaceTB/instruct-data-basics-smollm-H4: 1.0 
+
+dataset_splits:
+- train_sft
+- test_sft
+preprocessing_num_workers: 36
+
+# SFT trainer config
+bf16: true
+dataset_kwargs:
+  add_special_tokens: false  # We already wrap <bos> and <eos> in the chat template
+  append_concat_token: false # No need to add <eos> across samples
+do_eval: true
+evaluation_strategy: epoch
+gradient_accumulation_steps: 4
+gradient_checkpointing: true
+gradient_checkpointing_kwargs:
+  use_reentrant: false
+hub_model_id: smollm-360M-instruct-new
+hub_strategy: every_save
+learning_rate: 1.0e-03 # 3e-4
+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: data/smollm-360M-instruct-new
+overwrite_output_dir: true
+per_device_eval_batch_size: 4
+per_device_train_batch_size: 4
+push_to_hub: true
+remove_unused_columns: true
+report_to:
+- tensorboard
+- wandb
+save_strategy: "no"
+seed: 42
+warmup_ratio: 0.1