Hugging Face's logo

Datasets:
evalengine
/
imd-oracle-outcomes

Tasks:
Text Classification
Modalities:
Text
Formats:
json
Size:
10K - 100K
Tags:
imd
oracle
blockchain
disagreement-prediction
evaluation
Libraries:
Datasets
pandas
Polars
License:
Dataset card Data Studio Files
xet
 Community
1
imd-oracle-outcomes / README.md
johnsonchromia's picture
johnsonchromia
Add current model benchmarks and GPT-6.1 Sol results
9205cf6 verified
preview code
|
Raw History Blame Contribute Delete
4.84 kB
metadata 
license: unknown
pretty_name: IMD Oracle Outcomes
task_categories:
  - text-classification
tags:
  - imd
  - oracle
  - blockchain
  - disagreement-prediction
  - evaluation
size_categories:
  - 1K<n<10K
configs:
  - config_name: outcomes
    default: true
    data_files:
      - split: train
        path: data/train.jsonl
      - split: validation
        path: data/validation.jsonl
      - split: test
        path: data/test.jsonl
  - config_name: outcomes_2026_10_13
    data_files:
      - split: train
        path: snapshots/2026-10-13/data/train.jsonl
      - split: validation
        path: snapshots/2026-10-13/data/validation.jsonl
      - split: test
        path: snapshots/2026-10-13/data/test.jsonl

IMD Oracle Outcomes

Eval Engine's community dataset of 7,804 IMD oracle requests and 915,153 panel responses, collected from the IMD API. Includes questions, recorded answers, outcomes and source provenance.

Load 

from datasets import load_dataset
updated = load_dataset("evalengine/imd-oracle-outcomes", "outcomes_2026_10_13")

The current benchmark has 4,533 training, 1,349 validation and 381 test examples. Use input_json as model input; label is 0 for attested and 1 for disagreed. These labels describe protocol outcomes, not factual correctness.

The default outcomes configuration retains the original October 8 splits. Keep the two versions separate: their training and evaluation assignments differ.

Files and answers

  • Full archive: all requests, panel responses and provenance. Available answers appear in body.attestation.answer, body.agreement.answer and body.members[].answer.
  • Split metadata, audit and checksums: counts, cutoffs, deduplication and integrity checks.

Model benchmarks

Task: predict attested or disagreed from request fields only. These untuned baselines use the same frozen prompt and a 0.5 threshold. Precision and recall measure the disagreement class.

The current test has 381 examples, including 70 disagreements, from outcomes_2026_10_13.

Model Dataset used Precision Recall Average precision
GLM 5.3 Flash Current, test 47.67% 58.57% 0.5253
GLM 5.3 Current, test 42.22% 81.43% 0.4273
Kimi K3 Current, test 47.83% 31.43% 0.4485
GPT-6 Luna Current, test 41.00% 58.57% 0.3238
GPT-6.1 Sol Current, test 60.44% 78.57% 0.6037

Sol completed the test. Validation was deferred after gateway errors, with saved but unscored predictions: 140/1349 validation.

The current runs requested low reasoning and a 4,096-token output limit. Sol and Luna ran through a local gateway; the GLMs and Kimi used OpenRouter. Providers and quantization differ, so this compares the recorded serving setups. The gateway did not report dollar costs. Full results, usage and settings.

Current validation results

There are 1,349 validation examples but only six disagreements. One case changes recall by 16.7 percentage points. Keep validation and test scores separate: disagreement prevalence is 0.44% and 18.37%, respectively.

Model Dataset used Precision Recall Average precision
GLM 5.3 Flash Current, validation 2.13% 50.00% 0.0750
GLM 5.3 Current, validation 1.19% 66.67% 0.0172
Kimi K3 Current, validation 1.36% 50.00% 0.0137
GPT-6 Luna Current, validation 0.64% 50.00% 0.0073
GPT-6.1 Sol Current, validation; incomplete Not scored Not scored Not scored
Original dataset results

These results use the original 1,133 validation examples, including 18 disagreements. They are not directly comparable with the current splits. Original results and settings.

Model Dataset used Precision Recall Average precision
Always attested Original, Oct 8 0.00% 0.00% 0.0159
Qwen3.8-27B FP8 Original, Oct 8 33.33% 5.56% 0.0707
GLM 5.3 Flash Original, Oct 8 8.89% 22.22% 0.0587
GLM 5.3 Original, Oct 8 11.11% 33.33% 0.0720

The current validation and test sets have both been evaluated; no fine-tuning or outcome-driven prompt or threshold changes were used. The original test set remains unevaluated. Splits use creation time and retain one example per normalized question template. Recorded answers have not been independently verified; exclude them from outcome-prediction inputs.

License: unknown. Collected from the publicly accessible IMD API.