testnet-training-node-quickstart

This repository contains a demo script for you to fine-tune a Gemma model for train.flock.io.

Quickstart

Install Dependencies

To set up your environment, run the following commands:

conda create -n training-node python==3.10
conda activate training-node
pip install -r requirements.txt

File Structure

• dataset.py - Contains the logic to process the raw data from demo_data.jsonl.
• demo_data.jsonl - Follows the shareGPT format. The training data you receive from the fed-ledger is in exactly the same format.
• merge.py - Contains the utility function for merging LoRA weights. If you are training with LoRA, please ensure you merge the adapter before uploading to your Hugging Face repository.
• demo.py - A training script that implements LoRA fine-tuning for a Gemma-2B model.
• full_automation.py - A script that automate everything including get a task, download the training data, finetune Gemma-2B on training data, merge weights, upload to your HuggingFace model repo, and submit the task to fed-ledger.
• training_args.yaml - A YAML defines the training hyper-parameters for fine-tuning. A detailed explanation on LoRA config can be found here: LoRA Fine-tuning & Hyperparameters Explained

Full Automation

Simply run

TASK_ID=<task-id> FLOCK_API_KEY="<your-flock-api-key-stakes-as-node-for-the-task>" HF_TOKEN="<your-hf-token>" CUDA_VISIBLE_DEVICES=0 HF_USERNAME="your-hf-user-name" python full_automation.py

The above command will automatically train and submit multiple LLMs that are smaller than the max parameters limitation for the given task. Ensure that each model version is committed to the same repository to maintain version history. Refer to the Hugging Face documentation on commits for guidance on how to manage commits effectively.

Bypass certain models

If you want to bypass certain models, simply comment out the model config in the training_args.yaml

---

Play with demo.py

Start the Training

Execute the following command to start the training:

HF_TOKEN="hf_yourhftoken" CUDA_VISIBLE_DEVICES=0 python demo.py

The HF token is required due to the Gemma License.

This command initiates fine-tuning on the demo dataset, saves the fine-tuned model, merges the adapter to the base model, and saves the final model.

Upload the model folder to your HuggingFace repo

To maintain a clean version history and simplify change tracking, commit your model changes to the existing repository rather than creating a new one for each base model. For comprehensive guidance on uploading models and managing commits, consult the Hugging Face documentation on models uploading and the guide on commits.

Getting the task id

Before you submit the model script for a task, you will first need to stake on the task as a node.

• Go to the FLock Training Platform
• Select Training Node tab and stake on the task you want to submit the model to.
• The task ID is the ID of the task you want to submit the model to.

Submit the model

curl --location 'https://fed-ledger-prod.flock.io/api/v1/tasks/submit-result' \
--header 'flock-api-key: <your-api-key-with-staking-on-this-task-as-node>' \
--header 'Content-Type: application/json' \
--data '{
    "task_id": 29,
    "data":{
        "hg_repo_id": "Qwen/Qwen1.5-1.8B-Chat",
        "base_model": "qwen1.5",
        "gpu_type": "<GPU-used-for-training>",
        "revision": "<huggingface-model-commit-hash>"
    }
}'

• task_id: The ID of the task to which you are submitting a model.
• hg_repo_id: The Hugging Face repository where the model is stored, typically in the format username/repository-name.
• base_model: The base model used for training. A list of all supported models can be found here.
• gpu_type: The type of GPU used for training the model.
• revision: The commit hash from the Hugging Face repository. This uniquely identifies the version of the model that was trained and submitted, allowing for precise tracking of changes and updates.