TextInsightBench / docs /SUBMISSIONS.md
erwinmsmith's picture
Rebuild open exploration tasks and corpus-grounded evaluation
149da17 verified
|
Raw
History Blame Contribute Delete
4.79 kB

Submission contract

Write one UTF-8 JSON file per task, named <task_id>.json. The agent process protocol writes that same object to stdout. Do not wrap it in Markdown.

{
  "task_id": "the task ID",
  "findings": [],
  "abstention_reason": "No sufficiently supported finding was identified."
}

A nonempty answer contains up to three findings. Every finding has:

Field Required content
finding_id Unique identifier within the answer
claim Specific downstream conclusion, including direction and scope
kind Exactly the task's kind
scope Corpus, group, time and entity scope
population Object with filters: zero to three conjunctive metadata rules
comparison Agent-selected groups/date boundary, or null for association
definitions One observable condition for group/time tasks; two for association tasks. Each has condition_id, inclusion, exclusion
assignments One complete document partition per condition: condition_id, positive_doc_ids, negative_doc_ids, unknown_doc_ids
statistics Exactly the recomputed fields below
evidence Original quotation records: doc_id, start, end, quote, role (supporting, counterexample, context)
limitations Nonempty list describing uncertainty, confounding and inference limits

At least three distinct supporting documents and a known negative/discordant counterexample (if assigned cases exist) are required, with at most 15 spans per finding. Every selected-population document belongs to exactly one state per condition. Positive means the stated report is present; negative means it is not reported under the definition; unknown preserves unresolved judgments. Quotations refer to the text field. Python text[start:end] must equal quote, using Unicode characters, not UTF-8 bytes or JavaScript UTF-16 code units. Do not normalize or edit evidence text before computing offsets.

Population example: {"filters":[{"field":"rating","op":"gte","value":2}]}. Use {"filters":[]} for all documents. Allowed fields are listed in each task; operators are eq, in, gte and lte. An in list has 1–20 scalar values. Missing metadata never matches a filter. Document IDs, text and post-hoc condition labels cannot filter the population. Explain scope choices in scope and limitations.

Group example: {"field":"rating","groups":[[1,2],[4,5]]}. Each group has 1–20 distinct values, and the groups must be disjoint. This example is a syntax illustration, not a sufficient research finding. Group values omitted from both arms remain in the selected population and require assignments; statistics report their exclusion. Temporal example: {"field":"timestamp","cutoff":"2023-06-01"}. Compound association uses null. Minimum population and arm sizes are specified in each task.

Use the implementation to compute statistics from your assignments:

from textinsightbench.validation import expected
finding["statistics"] = expected(finding, task, documents)

For group/time tasks, required statistics are group0_total_n, group0_known_n, group0_positive_n, group0_unknown_n, group0_rate_known and the corresponding five group1_* fields, plus excluded_metadata_n, delta_known_pp, delta_identification_lower_pp, delta_identification_upper_pp.

Group order follows comparison.groups; for time tasks, group 0 is before the cutoff and group 1 is on or after. The known rate is positive / (total − unknown). The reported difference is group 1 minus group 0, in percentage points. The lower/upper identification bounds allocate unknowns to all compatible states within the finite corpus. These bounds are not confidence intervals. Missing comparison metadata is excluded from named denominators but still needs a condition judgment.

For association tasks, required statistics are known_joint_n, unknown_joint_n, n11, n10, n01, n00, p_b_given_a, p_b_given_not_a, conditional_difference_pp, lift. Condition A is the first definition and B the second. Joint cells count documents with known judgments for both conditions; unknowns are counted separately. Zero denominators produce JSON null, never NaN or infinity. Additional unsupported statistics are not accepted by this contract.

Current findings also require corpus_total_n, population_total_n, population_coverage and all robustness_* fields returned by expected. Pass the entire task corpus to expected; it applies the declared population itself. Do not pre-filter twice. See audit formulas. The schema is bundled in textinsightbench/output.schema.json and the dataset root. Synthetic discovery fixtures appear in tests/test_discovery.py; they are not released-task answers.