Task Details

In Track A, your system gets a triple consisting of an anchor story and two choices: A and B. Your system is asked to identify which of the two choices is more similar.

Track B instead asks you to produce a vector representation for an individual story. These representations should have a cosine similarity that aligns with the underlying stories’ narrative similarities. We, the organizers, will validate your submitted representations against triple-wise similarity ratings.

The task is currently running here on CodaBench.

Baselines

We provide baseline systems for both tracks for you to build upon. The output format of the baselines is suitable for uploading to CodaBench. The baselines in the CodaBench leaderboard are the GPT-4o-mini prompting approach for Track A and a sentenceBERT approach (specifically “all-MiniLM-L6-v2”) for Track B.

Rules

In Track B, embeddings are, at inference time, to be created on individual story instances only. You may, for example, not create embeddings such that they align with Track A-style answers for a full cross-product of triples across the provided dataset.

Embeddings for Task B may be anywhere from 10 to 8192 in length.

Be kind, have fun!

How To Submit to CodaBench

You will need to upload a zip file with the following structure:

.
├── track_a.jsonl
└── track_b.{npy,jsonl}

Make sure to fulfill all these requirements:

  • Have a file for each track you are participating in (you can deselect either one on the upload page)
  • Each file needs to be at the root level of the zip file
  • The order in the files must correspond to the order in the provided data
  • For Track A, your labels need to be in the “text_a_is_closer” property of each line
  • For Track B, you may either place your embeddings as a list of floats in the embeddings property or provide a numpy serialized file.

The baseline systems showcase a possible way to create the data.

In the upload form, make sure to specify:

  1. the type of method your system is using in each track.
  2. which of the tracks (either one or both) you are submitting to.

Make sure to select the methodology and correct track