README.md

metadata

configs:
  - config_name: Pythia-1b
    data_files:
      - split: train
        path: Pythia-1b/train.jsonl
      - split: ref
        path: Pythia-1b/ref.jsonl
  - config_name: Llama-3.2-1B
    data_files:
      - split: train
        path: Llama-3.2-1B/train.jsonl
      - split: ref
        path: Llama-3.2-1B/ref.jsonl
  - config_name: Llama-3.1-8B
    data_files:
      - split: train
        path: Llama-3.1-8B/train.jsonl
      - split: ref
        path: Llama-3.1-8B/ref.jsonl

Overview

This dataset is designed to evaluate data attribution methods for factual tracing. For each example in the reference set, there exists a subset of supporting training examples that we aim to retrieve.

Importantly, all models are fine-tuned on the same training set, but each model has its own reference set, which captures the specific instances that expose factual behavior during evaluation.

Structure

Each entry in the dataset contains the following fields:

• data_id (str): unique identifier.
• prompt (str): input query.
• response (str): training label.
• facts (List): List of strings containing the atomic facts it supports.

---

Stats

Model/Split	Train	Ref
Pythia-1b	6708	30
Llama-3.2-1B	6708	101
Llama-3.1-8B	6708	30

---

Example

{
  "data_id": "ftrace_0",
  "prompt": "Complete the sentence by filling in the blank:\n    Tamazight and other Berber varieties are spoken in Morocco, <blank>, Libya, Tunisia, northern Mali, and northern Niger by about 25 to 35 million people.\n    ",
  "response": "Algeria",
  "facts": ["P47,Q262,Q1028", "P37,Q25448,Q262", "P47,Q1028,Q262", "P47,Q1016,Q262", "P47,Q948,Q262", "P47,Q912,Q262", "P47,Q1032,Q262", "P47,Q262,Q1016", "P47,Q948,Q1016", "P47,Q1032,Q1016", "P47,Q262,Q948", "P47,Q1016,Q948", "P47,Q262,Q912", "P47,Q1032,Q912", "P47,Q262,Q1032"],
}