CLI usage¶
The command-line interface is extractinate. It runs a task file that already exists (build one in the Studio or by hand), which makes it ideal for repeatable and unattended runs.
Basic run:
extractinate --task_id 1 --model_name "phi4"
--task_id 1 selects Task001*.json from your task directory. On first use, Ollama pulls the model automatically.
The essentials¶
| Flag | Default | Purpose |
|---|---|---|
--task_id |
required | Which task to run (integer ID from the filename) |
--model_name |
phi4 |
Ollama model to use |
--run_name |
run |
Names the output subfolder — keep experiments apart |
--num_examples |
0 |
Turn on few-shot prompting (needs Example_Path in the task) |
--reasoning_model |
off | Required for "thinking" models (see below) |
--verbose |
off | Stream full logs — the first thing to reach for when debugging |
Directories default to data/, tasks/, examples/, and output/ in the current folder; override any of them (--data_dir, --task_dir, …) to point elsewhere. The settings reference documents every flag and every task-file field.
Reasoning models¶
Models that emit intermediate reasoning before their answer (DeepSeek-R1, Qwen3, and similar) need --reasoning_model so the tool pulls the final JSON out of the surrounding thought process:
extractinate --task_id 1 --model_name "deepseek-r1" --reasoning_model
These models also need more room to generate; that allowance is added for you, and if output still comes back empty or truncated you can raise --num_predict explicitly.
A few common recipes¶
Reproducible run (fixed seed, deterministic sampling):
extractinate --task_id 1 --model_name "phi4" --seed 42 --temperature 0.0
Few-shot with a custom embedding model:
extractinate --task_id 1 --num_examples 5 --embedding_model "mxbai-embed-large"
Large dataset, saved incrementally (crash-resistant):
extractinate --task_id 1 --model_name "phi4" --chunk_size 500
Widely varying document lengths (right-size the context per subset):
extractinate --task_id 1 --model_name "phi4" --max_context_len split
What you get¶
Results are written to output/<run_name>/<TaskName>-run<N>/nlp-predictions-dataset.json, with logs under output/logs/. Each record carries your input columns, the extracted fields, and a status. See Understanding output — and always spot-check.