Datasets:

sionic-ai
/

position-bias-training-datasets

license: mit
task_categories:
  - text-retrieval
language:
  - en
size_categories:
  - 10K<n<100K
configs:
  - config_name: begin
    data_files:
      - split: train
        path: begin/dist_100_0_0.json
  - config_name: middle
    data_files:
      - split: train
        path: middle/dist_0_100_0.json
  - config_name: end
    data_files:
      - split: train
        path: end/dist_0_0_100.json
  - config_name: uniform
    data_files:
      - split: train
        path: uniform/dist_33_33_33.json

Position-Bias Training Datasets (main 4 distributions)

Synthetic retrieval training data used to induce/control position bias in embedding models, restricted to the four main evidence-position distributions used in the paper. Filtered at reranker threshold t030 (signal_gap ≥ 0.30, highest label purity).

Each example is a (question, positive_doc) pair. The distributions differ only in where the relevant evidence sits inside the positive document (beginning / middle / end), mixed at the ratios below.

Configs (evidence-position distribution)

config	ratio (begin/middle/end)	evidence location	records
`begin`	100 / 0 / 0	evidence at document beginning	40,915
`middle`	0 / 100 / 0	evidence in document middle	40,915
`end`	0 / 0 / 100	evidence at document end	40,915
`uniform`	33.3 / 33.3 / 33.3	evidence uniformly across positions	40,915

Schema

question (string): synthetic query (apply the model's query prompt at encoding time).
positive_doc (string): the relevant document containing the evidence.

Usage

from datasets import load_dataset
ds = load_dataset("sionic-ai/position-bias-training-datasets", "begin", split="train")

Provenance

Subset of the private dataset sionic-ai/position-bias-train-030 (13 distributions); this repo keeps only the 4 main distributions with semantic config names (begin=100_0_0, middle=0_100_0, end=0_0_100, uniform=33_33_33).