FinEredium1/Forecast-Dojo

Forecast Dojo

Replayable forecasting environments for benchmarking and training LLM agents. This snapshot is prepared from the supplied Forecast-Dojo_results.zip and matches the scope of paper v1. The archive does not declare an export version or model sampling configuration; source_release=paper-v1 is this export's descriptive provenance label.

Snapshot

• Training: 1,338 events and 5,325 dated forecast steps.
• Evaluation: 230 events and 797 dated forecast steps.
• Total: 1,568 events and 6,122 dated forecast steps.
• 12 models, 36 conditions, four repeats per condition.
• 114,746 of 114,768 expected checkpoint-repeat forecasts.
• 22 forecasts missing: four GPT-5.6 Sol memory-on and 18 Nemotron memory-on.
• 37,134 extracted submitted belief notebooks and 218,073 per-tool summaries.

Each model has three conditions: dated research with memory-free, dated research with memory-on, and memory-free without tools. Memory-on carries the submitted belief notebook between forecast steps. No-tool conditions contain no research tool calls and have retrieval=none; dated research has retrieval=baseline. Raw mode identifiers remain independent and sequential for schema compatibility.

Data tables

• questions: question text, resolution criteria, outcomes, dates, domains, market metadata, resolved labels, and historical crowd probabilities.
• forecasts: submitted distributions, recorded scores, validation diagnostics, token counts, model latency, tool totals, and costs.
• notebooks: the final submitted <belief_notebook> block, extracted from the response. These are belief state records, not model reasoning transcripts.
• tool_usage: per-checkpoint aggregate calls, successes, errors and latency.
• runs: condition metadata, observed coverage and hashes of the source JSONL files.

Scoring and provenance

• Recorded accuracy and Brier scores are preserved, including recorded failure scores.
• Missing forecasts stay missing. They are never imputed or scored as zero.
• Eight evaluation checkpoints have no crowd probability; their Information alpha is null in the raw logs. Paper v1 scores normalize usable distributions and replace unusable recorded forecasts with uniform probabilities for all metrics, including Information alpha when the scalar market reference is available. Raw logged scores and parser diagnostics remain unchanged in the Parquet tables.
• parse_ok concerns answer extraction. probability_vector_valid and labels_match_question separately describe the submitted distribution.
• checkpoint_index is the chronological position in the question schedule. episode_step_index is the runner-local step and resets for memory-free runs.
• Notebook extraction preserves invalid JSON as text and records json_valid, block_count and the source formatting diagnostic.
• dataset_manifest.json records the ZIP hash, exact source members and coverage.
• Training event 38374 contains two source markets named Hungary. Both market records are retained; options lists each outcome label once. The manifest records this source warning. One training event has domain unknown, preserved.
• The ZIP contains no updated precomputed paper-analysis files. Old analysis outputs are retired rather than presented as results of this snapshot.
• Browser-ready assets for this snapshot are under web/paper-v1-cd98a865/. Their chart and trajectory scores follow paper v1, Appendix B.3/B.4, rather than the raw log failure penalties. Older web/ paths are retained for previous website versions. The website uses this versioned path so different releases cannot be mixed.

This export excludes full unparsed model responses, system prompts, internal endpoint configuration and the CC-News article corpus. Earlier dataset versions remain available through the repository commit history.