BTZSC

A unified benchmark for zero-shot text classification across embedding models, cross-encoders, rerankers, and LLMs.

Installation | Usage | Leaderboard | Documentation | Citing

---

Overview

BTZSC is a benchmark package for evaluating zero-shot text classification models under a unified interface. It helps you compare very different model families using the same datasets, task groupings, and metrics.

It is also the evaluation harness behind the BTZSC Hugging Face leaderboard: you can run the benchmark locally, export a leaderboard-ready JSON artifact, and submit new entries to keep the public results up to date.

The package includes:

• Dataset loaders for BTZSC benchmark tasks.
• A shared benchmark runner across model adapters.
• Built-in adapters for embedding, NLI, reranker, and LLM-style models.
• Baseline comparison utilities and a CLI for reproducible evaluation.

Paper and Resources

• Paper (OpenReview): https://openreview.net/forum?id=IxMryAz2p3
• PDF: https://openreview.net/pdf?id=IxMryAz2p3
• Eval harness (GitHub): https://github.com/IliasAarab/btzsc
• Leaderboard results dataset: https://huggingface.co/datasets/btzsc/btzsc-results
• Leaderboard Space: https://huggingface.co/spaces/btzsc/btzsc-leaderboard

For Users

Installation

Install with pip:

pip install btzsc

Install with uv in an existing project:

uv add btzsc

Run as a standalone CLI tool with uvx (no project install needed):

uvx btzsc list-datasets

Quick Start (Python API)

Use this as a recommended first workflow:

1. Start with one or two task groups to validate your setup.
2. Inspect summary and per-dataset outputs.
3. Compare against bundled baselines.
4. Export a leaderboard-ready JSON artifact.

API notes:

• BTZSCBenchmark(tasks=...) accepts either task groups ("sentiment", "topic", "intent", "emotion") or explicit dataset names. Leave empty to run all datasets.
• evaluate(model=..., model_type=...) returns a BTZSCResults object.
• model_type is required when model is a string model ID (if you pass a BaseModel instance, you can omit it). Choose from embedding, nli, reranker, llm.
• Use max_samples for quick smoke tests; increase batch_size for throughput if your hardware allows it.

from btzsc import BTZSCBenchmark

benchmark = BTZSCBenchmark(tasks=["sentiment", "topic"])
results = benchmark.evaluate(
	model="intfloat/e5-base-v2",
	model_type="embedding",
	batch_size=64,
)

print(results.summary())
print(results.per_dataset())

# Compare against bundled baselines
print(results.compare_baselines(metric="f1"))

# Export leaderboard-ready JSON
results.to_json("results/embedding/e5-base-v2.json")

Quick Start (CLI)

Equivalent end-to-end CLI flow:

Note: when --model is a model ID string, you must also provide --type.

# 1) Explore benchmark metadata
btzsc list-datasets
btzsc list-model-types

# 2) Run an initial benchmark
btzsc evaluate --model intfloat/e5-base-v2 --type embedding --tasks sentiment,topic

# 3) Compare with packaged baselines
btzsc baselines --metric f1 --top 10

# 4) Export JSON for leaderboard submission
btzsc evaluate \
	--model intfloat/e5-base-v2 \
	--type embedding \
	--output-json results/embedding/e5-base-v2.json

# 5) Validate the JSON locally
btzsc validate-result results/embedding/e5-base-v2.json

Tip: run a small pilot first, then repeat with your full task scope for final reporting.

Supported Model Types

BTZSC currently supports these adapter families:

• embedding
• nli
• reranker
• llm

Pass the model type explicitly (model_type in Python or --type in CLI).

Extending with Custom Models

To make a custom model compatible with BTZSC, implement an adapter that subclasses BaseModel.

Contract requirements:

• predict_scores(texts, labels, batch_size) must return a score matrix with shape (len(texts), len(labels)) where higher means more likely.
• predict(texts, labels, batch_size) must return predicted label indices with shape (len(texts),).
• Set model_type on your class. Use embedding, nli, reranker, or llm when applicable.

import numpy as np

from btzsc.models.base import BaseModel


class MyCustomAdapter(BaseModel):
	model_type = "embedding"

	def __init__(self, model_name: str = "my-org/my-model"):
		self.model_name = model_name

	def predict_scores(
		self,
		texts: list[str],
		labels: list[str],
		batch_size: int = 32,
	) -> np.ndarray:
		# Replace this with your real scoring implementation.
		return np.zeros((len(texts), len(labels)), dtype=float)

	def predict(
		self,
		texts: list[str],
		labels: list[str],
		batch_size: int = 32,
	) -> np.ndarray:
		scores = self.predict_scores(texts, labels, batch_size=batch_size)
		return scores.argmax(axis=1)

Run it in the benchmark:

from btzsc import BTZSCBenchmark

benchmark = BTZSCBenchmark(tasks=["sentiment", "topic"])
custom_model = MyCustomAdapter("my-org/my-model")

results = benchmark.evaluate(
	model=custom_model,
	batch_size=32,
	max_samples=200,
)

print(results.summary())
results.to_json("results/custom/my-model.json")

When you pass a BaseModel instance to evaluate(), you do not need model_type=... in the call.

Submitting to the Leaderboard

After exporting your JSON (results.to_json(...) or --output-json), first validate it:

btzsc validate-result results/<model_type>/<model-name>.json

Then publish it to the results dataset repo:

https://huggingface.co/datasets/btzsc/btzsc-results

Required destination path format:

results/<model_type>/<model-name>.json

Example:

results/embedding/e5-base-v2.json

You can submit using any of these workflows:

1. Web UI (no clone required)

   • Open the results repo page: https://huggingface.co/datasets/btzsc/btzsc-results
   • Go to Files and versions and upload your JSON at the required path.
   • If you do not have write access, fork the repo and open a PR.

2. Git workflow (clone/fork + push)

   • Clone (or fork) btzsc/btzsc-results, add your JSON at the required path, then push.
   • If you pushed to a fork, open a PR to btzsc/btzsc-results.

git lfs install
git clone https://huggingface.co/datasets/btzsc/btzsc-results
cd btzsc-results

# Copy your exported JSON into the correct folder
mkdir -p results/reranker
cp /path/to/my_result.json results/reranker/my-model.json

git add results/reranker/my-model.json
git commit -m "Add BTZSC results for my-model"
git push

3. API workflow (huggingface_hub, PR-based)

   • Authenticate first (huggingface-cli login or HF_TOKEN).
   • create_pr=True creates a PR branch instead of pushing directly to main.

from huggingface_hub import HfApi

api = HfApi()
api.upload_file(
    path_or_fileobj="results/reranker/my-model.json",
    path_in_repo="results/reranker/my-model.json",
    repo_id="btzsc/btzsc-results",
    repo_type="dataset",
    commit_message="Add BTZSC results for my-model",
    create_pr=True,
)

The leaderboard Space reads from this results dataset and updates as new valid entries are added.

For full submission requirements, see hf/results_repo/SUBMISSION.md.

Benchmark Protocol

BTZSC follows a strict zero-shot protocol:

• 22 English single-label datasets
• 4 task families: sentiment, topic, intent, emotion
• No BTZSC-label training or tuning on evaluation datasets
• Primary leaderboard metric: macro-F1
• Secondary metrics: accuracy, macro-precision, macro-recall

The leaderboard is continuously updated as new submissions are added.

Dataset

BTZSC benchmark data is available on Hugging Face:

https://huggingface.co/datasets/btzsc/btzsc

To load the raw paired-format rows with datasets:

from datasets import get_dataset_config_names, load_dataset

repo_id = "btzsc/btzsc"

# Each dataset is a config name (e.g. "agnews", "imdb", ...)
print(get_dataset_config_names(repo_id)[:5])

# Load one dataset's test split
ds = load_dataset(repo_id, "agnews", split="test")
print(ds.column_names)
print(ds[0])

The dataset stores rows as (text, hypothesis, labels) where labels is binary entailment. The package reconstructs grouped multiclass samples internally for evaluation.

Citing

@inproceedings{aarab2026btzsc,
	title     = {BTZSC: A Benchmark for Zero-Shot Text Classification Across Cross-Encoders, Embedding Models, and Rerankers},
	author    = {Aarab, Ilias},
	booktitle = {International Conference on Learning Representations (ICLR) 2026},
	year      = {2026},
	note      = {OpenReview PDF: https://openreview.net/pdf?id=IxMryAz2p3},
	url       = {https://openreview.net/forum?id=IxMryAz2p3}
}

License

Released under the MIT license.

---

For Developers

Developer Setup

git clone https://github.com/IliasAarab/btzsc.git
cd btzsc
uv sync --dev

Project Structure

High-level layout:

• src/btzsc/benchmark.py: benchmark orchestration and result objects.
• src/btzsc/data.py: dataset loading and task grouping.
• src/btzsc/metrics.py: metric computation and summaries.
• src/btzsc/baselines.py: baseline loading and comparison table creation.
• src/btzsc/models/: model adapters (embedding, nli, reranker, llm).
• src/btzsc/cli.py: command-line interface.

Quality Checks

Run formatting, linting, and typing checks before opening a PR:

uv run ruff format
uv run ruff check
uv run pyright

Packaging and Release

Build locally:

uv build

Release process:

1. Bump version in pyproject.toml.
2. Commit and push to main.
3. Create and push a version tag, for example:

git tag v0.1.1
git push origin v0.1.1

GitHub Actions builds and publishes tagged releases to PyPI via trusted publishing.