README.md

metadata

license: cc-by-sa-4.0
task_categories:
  - question-answering
language:
  - en
pretty_name: parser_dataset
tags:
  - multi-hop-qa
  - hotpotqa
  - 2wikimultihopqa
  - retrieval
configs:
  - config_name: train
    data_files:
      - split: train
        path: train/hotpotqa_train_process_emb-select_llm-unable.parquet
  - config_name: 2wiki_val
    data_files:
      - split: eval_50
        path: 2wiki_val/eval_50.json
      - split: eval_100
        path: 2wiki_val/eval_100.json
      - split: eval_200
        path: 2wiki_val/eval_200.json
      - split: eval_400
        path: 2wiki_val/eval_400.json
      - split: eval_800
        path: 2wiki_val/eval_800.json
      - split: eval_1600
        path: 2wiki_val/eval_1600.json
      - split: eval_3200
        path: 2wiki_val/eval_3200.json
      - split: eval_6400
        path: 2wiki_val/eval_6400.json
  - config_name: hqa_val
    data_files:
      - split: eval_50
        path: hqa_val/eval_50.json
      - split: eval_100
        path: hqa_val/eval_100.json
      - split: eval_200
        path: hqa_val/eval_200.json
      - split: eval_400
        path: hqa_val/eval_400.json
      - split: eval_800
        path: hqa_val/eval_800.json
      - split: eval_1600
        path: hqa_val/eval_1600.json
      - split: eval_3200
        path: hqa_val/eval_3200.json
      - split: eval_6400
        path: hqa_val/eval_6400.json

parser_dataset

Parser training and evaluation data for multi-hop QA with long concatenated document contexts.

Derived from HotpotQA and 2WikiMultihopQA.

Contents

Path	Split	Samples	Notes
train/hotpotqa_train_process_emb-select_llm-unable.parquet	train	HotpotQA processed train	VERL / ParserRLHFDataset format
2wiki_val/eval_{N}.json	eval	128 per file	2WikiMultihopQA, N documents per example
hqa_val/eval_{N}.json	eval	128 per file	HotpotQA, N documents per example

N ∈ {50, 100, 200, 400, 800, 1600, 3200, 6400}.

Load

from datasets import load_dataset

train = load_dataset("inNexus/parser_dataset", "train", split="train")
hqa_50 = load_dataset("inNexus/parser_dataset", "hqa_val", split="eval_50")
wiki_6400 = load_dataset("inNexus/parser_dataset", "2wiki_val", split="eval_6400")

Fields

Train parquet (consumed by ParserRLHFDataset): prompt, context, reward_model, data_source.

Eval JSON

• hqa_val: context, input, answers, num_docs, index
• 2wiki_val: same as above, plus is_ordered, is_reverse_ordered, min_distance, evidence_idx, level, type, id