aurorajy's picture
Limit experiment release to report studies and documented selection provenance
2635280 verified
|
Raw History Blame Contribute Delete
3.46 kB
metadata
language:
  - en
task_categories:
  - question-answering
tags:
  - turtle-soup
  - lateral-thinking
  - multi-agent
  - in-context-learning
pretty_name: GoodTurtle Experiments
configs:
  - config_name: games
    data_files:
      - split: records
        path: games.jsonl
  - config_name: turns
    data_files:
      - split: records
        path: turns.jsonl
  - config_name: messages
    data_files:
      - split: records
        path: messages.jsonl

GoodTurtle experiment records

Experiment evidence accompanying GoodTurtle. Puzzle collection: GoodTurtle dataset.

Report scope

This version removes 61 unused legacy games and adds 18 completed supplementary games used in the report. It contains 163 games across CL, the three main MAS targets, the reflection pilot and paired output-budget controls. The fourth MAS-reflection interruption is replaced by its completed continuation.

The CL four-run descriptive subset is outcome-selected: positive groups use the highest four scores and baseline/negative/irrelevant groups use the lowest four (stable order resolves ties). Complete candidate records and results/report_provenance/cl_selection.csv are retained in the archive so the selection can be audited. The negative/irrelevant full-dialogue appendix likewise uses its lowest four scored games. These are selected descriptive comparisons, not randomly selected four-run samples.

The 50-puzzle dataset is unchanged. The original game wording, host replies and scores are preserved.

Contents

  • games: one row per canonical archived game, with condition, outcome, model and usage information.
  • turns: original questions and submitted explanations with recorded public feedback.
  • messages: shared agent messages and private draft/review notes, explicitly distinguished.
  • results.tar.gz: full published text/JSON records, frozen settings, review records and analysis tables.
  • archive_manifest.json: inventory for checking the complete downloaded archive.
  • release.json: counts, source code version and archive information.

Use game_id to join the tables. The records split includes source games and technical interruptions; filter role, status and condition explicitly for statistical comparisons. Recovery copies are not additional trials.

Provenance and interpretation

Original questions, host replies, scores, source labels and completion decisions are unchanged. Model-generated gameplay and private reflection remain model-generated records, distinct from human reference annotations. Historical cohorts retain their original model, host and grading versions. Revised report labels for Set 2 are documented in the code repository; original labels are not retroactively rewritten here.

English is the primary language. Original messages and dated review notes may retain other languages. This archive contains model judgments and research annotations; it is not a human-gold benchmark of every reply. Potentially violent fictional stories appear in the source puzzles and gameplay.

Reproduce and cite

Download a fixed dataset revision and use the code repository's scripts/fetch_results.py to restore and verify the complete archive before running analysis. See the repository's CITATION.cff for author metadata. Underlying puzzle stories retain their source ownership; this dataset does not assign a new license to them.