Labrador Synthetic Data Generation Pipeline
Introduction
This repository contains the Labrador synthetic data generation pipeline, which is used to generate synthetic data for various purposes.
Run Instructions (Automation)
Step 1: Environment Setup
- Initialize a
.env
file with the following access tokens:GIT_ACCESS_TOKEN={ACCESS-TOKEN-TO-ACCESS-TAXONOMY-REPO} # this personal access token is used to access instruct-lab/taxonomy repo
Step 2: Execution
To run the pipeline:
Execute the following command:
NOTE: Depending on whether you are running on old or new vela, change this line in the orchestrator.py to use the appropriate old vela or new vela template.
save_job_with_jinja_template(cfg, "templates/labrador_datagen_vela.yaml.j2", output_dir=f"jobs/{branch}")
python orchestrator.py branch-name
This will:
- Create a file with a list of leaf nodes in the
jobs
directory. - Generate YAML files for each leaf node and store them in the
jobs
directory something liketest-7984f9cae729b798bed1ba222715b880.yaml
- Create a file with a list of leaf nodes in the
To initiate the skill generation pipeline, run:
To trigger a job, take the above yaml and
oc create -f jobs/yaml_name.yaml
This command will execute the pipeline and store the results in the
new_data/labrador-datagen
directory within the COS bucket mounted on the Vela cluster.
Run Instructions (Manual - Testing)
Step 1: Run model
Run teacher model - this model can be replaced with any small model for testing purposes
text-generation-launcher -p 8080 --model-id mistralai/Mixtral-8x7B-Instruct-v0.1 --dtype bfloat16 --max-input-length 4096 --max-batch-prefill-tokens 4096 --max-total-tokens 12288
Next, set the following enviornment variables:
LEAF_NODE=knowledge/textbooks/ethics/qna.yaml # Path to the leaf node that you want to download
NUM_SAMPLES=30
NUM_GROUNDED_QUESTIONS=3
NUM_GEN_PROC=32
NUM_UTIL_PROC=8
SAVE_PATH=new_data/labrador_datagen # Path where you want to download the data
CONTEXT=0 # Set 0 for freeform and 1 for grounded
DATA_PATH=.
CHECKSUM=test
BRANCH_NAME=test # Branch name to download data from
KNOWLEDGE=1 # Set 0 for skills and 1 for knowledge
PARENT_DIR=$(dirname "$LEAF_NODE")
GIT_ACCESS_TOKEN= # Access token to access taxonomy repo
Skills
Download data
wget --header "Authorization: token $GIT_ACCESS_TOKEN" --directory-prefix="$DATA_PATH/$PARENT_DIR" "https://raw.githubusercontent.com/instruct-lab/taxonomy/$BRANCH_NAME/$LEAF_NODE"
Run the Justfile using:
just run
The Justfile will check the context value. If the context is set to 1, it will run scripts for grounded data generation. If the context is set to 0, it will run scripts for freeform data generation and save the generated files in the root of the repo in the same directory structure.
Knowledge
Download data
bash download_docs.sh
Run knowledge script
python knowledge_generation_pipeline.py
- Downloads last month
- 7