llm-annotator 0.7.2

An easy-to-extend LLM annotator for robust, resumable data annotation.

Robust, resumable LLM dataset annotation

llm-annotator is a Python 3.12+ library for robust, resumable LLM-driven dataset annotation and generation.

It supports multiple providers through pluggable clients:

• vLLM offline inference: VLLMOfflineClient
• vLLM server API: VLLMClient
• OpenAI API: OpenAIClient
• Anthropic API: ClaudeClient
• Gemini API: GeminiClient

Key capabilities:

• Resumable processing with JSONL checkpoints.
• Annotation of existing datasets and generation from scratch.
• Structured outputs via JSON schema.
• Retry and validation hooks for robust pipelines.
• Optional Hugging Face Hub upload cadence.
• Context-manager cleanup of client resources.

Documentation

Read the full documentation at bramvanroy.github.io/llm-annotator.

Provider setup reference: docs/provider-info.md

Installation

Recommended:

uv add llm-annotator

or

pip install llm-annotator

Install provider extras as needed:

uv add "llm-annotator[vllm]"
uv add "llm-annotator[openai]"
uv add "llm-annotator[anthropic]"
uv add "llm-annotator[gemini]"

See docs/provider-info.md for auth environment variables and provider-specific setup notes.

For local vLLM runs, install flashinfer for your CUDA version.

uv pip install flashinfer-python flashinfer-cubin
# JIT cache package (replace cu128 with your CUDA variant)
uv pip install flashinfer-jit-cache --index-url https://flashinfer.ai/whl/cu128

Usage

Annotate an existing dataset:

from llm_annotator import Annotator, VLLMOfflineClient

# Use a local vLLM model
client = VLLMOfflineClient(
    model="meta-llama/Llama-3.2-3B-Instruct",
    max_model_len=4096,
)

with Annotator(client=client, verbose=True) as anno:
    ds = anno.annotate_dataset(
        output_dir="outputs/sentiment",
        prompt_template="Classify the sentiment of this text: {text}",
        dataset_name="stanfordnlp/imdb",
        dataset_split="test",
        max_num_samples=100,
    )

Generate a dataset from scratch:

from llm_annotator import Annotator, OpenAIClient

client = OpenAIClient(model="gpt-4o-mini")

with Annotator(client=client) as anno:
    ds = anno.generate_dataset(
        output_dir="outputs/generated-qa",
        prompts="Write a short geography quiz question with answer.",
        max_num_samples=200,
    )

See the documentation for more examples, including:

• Structured output with JSON schemas
• Custom validation and post-processing
• Large-scale streaming annotation
• Generating datasets from scratch
• Multi-GPU support

Or check out the examples/ directory for complete working examples.

Testing

Install development dependencies first:

uv sync --dev

Run the default checks:

make style
make quality
make test
make typecheck

Pytest marker targets:

# Fast tests (same as `make test`)
make test-fast

# Slow tests only
make test-slow

# Integration tests only
make test-integration

# Entire suite (fast + slow)
make test-all

You can also run markers directly with pytest:

uv run pytest -m "not slow"
uv run pytest -m "slow"
uv run pytest -m "integration"

Slow and integration tests may load local models, require more runtime, or depend on optional components.

Building documentation

Local versioned docs preview (uses mike on a temporary local branch):

make serve-docs

Override version metadata when needed:

make serve-docs DOCS_VERSION=0.4.0 DOCS_ALIAS=latest DOCS_SOURCE_REF=v0.4.0

Docs are published with mike on release tags through .github/workflows/docs.yml.