Shreyas94/needle-event-classifier-eval

Sentinel Needle Event Classifier — Evaluation Splits

Held-out test and validation splits used to evaluate the Shreyas94/needle-event-classifier model: a 13-class financial-news event classifier built on frozen Cactus Compute Needle 3 embeddings.

Important: this is NOT the training set

This repo contains only the frozen 678-row test split and the 424-row validation split, both human-reviewed. These are the exact splits the model's reported metrics (91.45% accuracy / 0.9057 macro F1 on the test split) were computed on.

The actual training data — 2,188 real, human-reviewed ("GOLD") rows plus 172 Gemini-augmented rows (2,360 rows total) — is not included here. It lives in a separate local dataset that was not available to push at the time this eval set was published. Do not assume these rows were part of training; they were explicitly held out and never used for fitting or hyperparameter tuning.

Files

• test.jsonl — 678 rows, {"headline": str, "true_label": str}.
• validation.jsonl — 424 rows, {"headline": str, "true_label": str}.

Label note

The source files for these splits predate a taxonomy change: fundraising was originally a separate 14th class, later merged into capital_markets (see the model's own methodology notes). The reported model metrics (91.45% / 0.9057) were computed with that merge applied in-memory at eval time, never written back to the original files on disk. To keep this published copy consistent with the shipped 13-class model, that same remap (fundraising → capital_markets) has been applied here.

Taxonomy (13 classes)

capital_markets, default, downgrade, earnings, liquidity, ma, management_change, other, portfolio_company_event, rating_action, refinancing, regulatory, upgrade

Usage

from datasets import load_dataset

ds = load_dataset("Shreyas94/needle-event-classifier-eval", data_files={
    "test": "test.jsonl",
    "validation": "validation.jsonl",
})

Credits

Curated as part of the Sentinel News Tracker project. Embeddings for the paired model are produced by Cactus Compute's Needle 3 (not included/redistributed here).

License

MIT.