Create fine-tune-LLM-with-Autotrain.md

docs/sophia/data-science/fine-tune-LLM-with-Autotrain.md

@@ -0,0 +1,100 @@
+# Autotrain
+
+Autotrain, developed by Hugging Face, is a platform designed to simplify training cutting-edge models in various fields: NLP, LLM, CV... [read more](https://huggingface.co/docs/autotrain/main/en/tasks/llm_finetuning)
+
+## Create Python Virtual Environment for Autotrain
+
+Let's first create an environment for Autotrain:
+
+```bash
+mkdir -p venv_autotrain
+python -m venv venv_autotrain --system-site-packages
+source venv_autotrain/bin/activate
+pip3 install autotrain-advanced
+```
+
+**Note:** If Autotrain doesn't work properly, you may have to reinstall `nvidia-ml-py`.
+
+```bash
+pip3 uninstall nvidia-ml-py3 pynvml
+pip3 install --force-reinstall nvidia-ml-py==11.450.51
+```
+
+## Train Dataset Format
+
+The dataset should have a column "text" containing the data to be trained on. [Example](https://huggingface.co/datasets/timdettmers/openassistant-guanaco)
+
+## Config File for Fine-Tuning Local LLM
+
+Here is an example to create a config file for supervised fine-tuning purposes:
+
+```yaml
+task: llm-sft
+base_model: meta-llama/Meta-Llama-3.1-8B-Instruct
+project_name: Llama-3-1-FT
+log: wandb
+backend: local
+data:
+  path: Path/to/the/training/dataset/folder
+  train_split: train
+  valid_split: null
+  chat_template: null
+  column_mapping:
+    text_column: text
+params:
+  block_size: 1024
+  model_max_length: 8192
+  epochs: 800
+  batch_size: 2
+  lr: 1e-5
+  peft: true
+  quantization: null
+  target_modules: all-linear
+  padding: right
+  optimizer: paged_adamw_8bit
+  scheduler: cosine
+  gradient_accumulation: 8
+  mixed_precision: bf16
+hub:
+  username: ***
+  token: hf_***
+  push_to_hub: true
+```
+
+[More details](https://huggingface.co/docs/autotrain/en/config)
+
+## Run Autotrain to Fine-Tune Using the Config File
+
+```bash
+cd Path/to/save/the/adapter
+autotrain --config path/to/config.yaml
+```
+
+## Merge Adapters with Base Model to Create New Model
+
+Adapters need to be merged with the base model in order to run. You can use the code below:
+
+```python
+from transformers import AutoModelForCausalLM, AutoTokenizer, AutoConfig
+import torch
+from peft import PeftModel
+import os
+
+adapter = "path/to/saved/adapters/"
+model_name = "project-name-from-config-file"
+adapter_path = os.path.join(adapter, model_name)
+base_model_path = "meta-llama/Meta-Llama-3.1-8B-Instruct"
+target_model_path = "path/to/save/fine-tuned/models" + model_name
+
+config = AutoConfig.from_pretrained(base_model_path)
+base_model = AutoModelForCausalLM.from_pretrained(base_model_path)
+
+merged_model = PeftModel.from_pretrained(base_model, adapter_path)
+
+tokenizer = AutoTokenizer.from_pretrained(adapter_path, trust_remote_code=True)
+merged_model = merged_model.merge_and_unload()
+
+print("Saving target model...")
+merged_model.save_pretrained(target_model_path)
+tokenizer.save_pretrained(target_model_path)
+config.save_pretrained(target_model_path)