lang_table / README.md
talzoomanzoo's picture
Add PoC1 matched-input dataset and documentation
349b0f9 verified
|
Raw History Blame Contribute Delete
4.12 kB
metadata
language:
  - en
license: cc-by-sa-4.0
task_categories:
  - question-answering
size_categories:
  - n<1K
tags:
  - transition-analysis
  - proof-of-concept
configs:
  - config_name: default
    data_files:
      - split: test
        path: lang_table.json
    field: examples

Prose–table representation transitions

Ten adapted WikiTableQuestions examples for PoC1 textual-representation experiments. Each example has four matched prompts (40 prompts total), using English prose (A) and Markdown tables (B). All inputs are text; there are no images.

Conditions and structure

Condition letters specify the representation of segments S1 and S2, respectively:

Condition S1 S2
AA A A
AB A B
BA B A
BB B B

Each question requires information from both segments. Meanings, identifiers, and the English question remain fixed across conditions. The fixed English prose question can introduce an additional boundary after S2 in AB and BB.

lang_table.json preserves the original project file, including top-level setup metadata and an examples list. The dataset configuration exposes that list as the test split. This split name denotes evaluation examples; no train/test sampling or held-out benchmark construction is implied.

Each example includes id, task, segments, question, question_language, and inputs. Gold answers are in expected_answers, with comparison rules in answer_comparison. Additional source, table, and segment_facts fields preserve provenance and auditing information.

Loading

from datasets import load_dataset

examples = load_dataset("talzoomanzoo/lang_table", split="test")
prompt = examples[0]["inputs"]["AB"]

To preserve the complete JSON document and its setup metadata:

import json
from huggingface_hub import hf_hub_download

path = hf_hub_download(
    repo_id="talzoomanzoo/lang_table",
    filename="lang_table.json",
    repo_type="dataset",
)
with open(path, encoding="utf-8") as f:
    document = json.load(f)
prompt = document["examples"][0]["inputs"]["AB"]

Send only the selected inputs value to the model. Gold answers and other metadata are for evaluation and auditing, and must not be included in the prompt.

Intended use and limitations

These small examples support setup validation for inference, token-position activation capture, representation comparisons, and probing experiments. They are not a representative benchmark. Establish model-specific accuracy in AA and BB before interpreting transition effects. Keep all four conditions of an example together when constructing any later evaluation partitions.

Real-world facts may be recalled without using the supplied segments. S1 maps opaque record IDs to entities or years; S2 maps IDs to attributes. Record IDs are independent of rank, year, and answer. Row orders are independently shuffled between segments, then held fixed across conditions.

Source, modifications, and license

Adapted from WikiTableQuestions by Panupong Pasupat and Percy Liang, introduced in Compositional Semantic Parsing on Semi-Structured Tables (2015). The upstream repository distributes the dataset under CC BY-SA 4.0; this adaptation retains that license.

Changes include selecting rows and task-relevant columns, rewording questions, introducing opaque record IDs, splitting facts into two segments, shuffling their row orders, and rendering matched prose and Markdown tables. Nonbreaking spaces are normalized to ordinary spaces. Original question IDs, questions, answers, table paths, source URLs, and one-based row numbers remain in each example's provenance. These are adapted examples, not an unmodified WikiTableQuestions evaluation split.

The design follows the project's PoC1 matched-input construction described in tue-meeting.pdf; that document is not included in this repository.