imd-oracle-outcomes / README.md
johnsonchromia's picture
Add Qwen explicit-format and checker-retry development results; preserve v1
53274d6 verified
|
Raw History Blame Contribute Delete
7.64 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
  - config_name: oracle_worker_pilot_v1
    data_files:
      - split: test
        path: node-pilot-v1/test.jsonl
  - config_name: oracle_protocol_v1
    data_files:
      - split: test
        path: protocol-v1/test.jsonl

IMD Oracle Outcomes and Worker Pilot

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

Protocol compliance

204 cases from 102 source requests test submission formatting and controlled repair. The answer and recipe are supplied; this does not test factual accuracy or full node operation.

Revised prompt, development evaluation: after inspecting the original failures, we added an explicit JSON structure, a synthetic example and checker-guided corrections. The cases and grader are unchanged. These reused public tasks are not an untouched test set.

Qwen3.8-27B First attempt After up to 2 corrections Avg / p95 full task Tokens / success
Construction 102/102 102/102 6.58s / 9.37s 1,957
Repair 102/102 102/102 6.33s / 8.41s 2,256

0 tasks needed retries; 0 recovered. Total: 429,749 reported tokens, 204 API requests and 11.0 minutes wall time.

Settings: ModelScope, thinking off, temperature 0, 4,096 output tokens per attempt, maximum 3 attempts, concurrency 2. No constrained decoding. Time and tokens include all attempts. Feedback uses the official checker and visible inputs only. Dollar cost is unknown. The two tracks share source groups; keep their scores separate.

Methodology · Results and frozen prompt · All attempts · Reproduce

The data remains configuration oracle_protocol_v1. Send only input_json with the frozen prompt; reference_json is evaluator data.

Original prompt and results (v1)

204 derived cases from 102 source requests: construct a valid submission or repair a deliberately corrupted one. The result and recipe are supplied. Scoring checks the pinned IMD schema and preservation of those values; this does not test factual answers or full node operation.

Qwen3.8-27B, thinking off Passed Avg / p95 task time Tokens / success
Construction 0/102 (0.0%) 6.65s / 9.74s N/A
Repair 16/102 (15.7%) 6.91s / 10.24s 9,197

188 failures returned the recipe as quoted text instead of a JSON object. This is a result for one frozen prompt and serving setup, not a general Qwen capability rating.

Settings: ModelScope, thinking off, temperature 0, 4,096 output tokens, one attempt per case, concurrency 2. The prompt names required fields but provides no formal output schema or worked example. No JSON-schema constrained decoding or interactive retries. Tokens include failed attempts; dollar cost is unknown. Construction and repair variants share source groups; keep their scores separate.

Methodology · Results and frozen prompt · Per-case outputs · Reproduce

Load configuration oracle_protocol_v1; send only input_json to the model. reference_json is evaluator data. Prior pilot and disagreement results remain below.

Oracle worker pilot

Primary metric: verified tasks completed / all assigned tasks. Models must use the official scanner, check the pinned block hash, and submit a correct answer and recipe that pass the official checker. Errors, abstentions and timeouts count as failures.

Model Setting Verified tasks Avg / p95 task time Tokens / success
Claude Opus 5.5 medium 20/20 8.28s / 12.32s 4,448
GPT-6.1 Sol medium 20/20 9.30s / 11.74s 2,740
Qwen3.8-27B thinking off 13/20 15.22s / 18.88s 8,619

Qwen returned the correct number on 20/20 tasks; 7 failures were explained by omitting the required block-hash pin check. Numeric correctness alone is not the full success criterion. A grader correction accepted one valid Qwen event format; all saved outputs were rescored without rerunning models.

Scope: 20 simple Ethereum count/sum tasks; 19 already provide recipes. This is a pilot, not a general node ranking. Ground truth was recomputed from matching logs retrieved from two RPC providers. These tasks use an API tool agent and frozen RPC replay, not native IMD workers. Contract builds, audits, research, admission and earnings are untested.

Settings: temperature 0; up to 8 turns, 4,096 output tokens/turn and 180 seconds/task; concurrency 2. Opus/Sol use a local gateway; Qwen uses ModelScope. Time covers the full task; tokens include failed tasks. One run per model; provider tokenizers and reasoning settings differ. Dollar costs were not reported.

The worker's published premium requirements include Opus and Sol at medium effort. Machine/runtime admission still needs checking; Qwen is a research comparison, not a verified premium-worker configuration.

Methodology · Results and settings · Per-task traces · Reproduce

Load

from datasets import load_dataset
pilot = load_dataset("evalengine/imd-oracle-outcomes", "oracle_worker_pilot_v1")

Supply only input_json to the model. reference_answer is evaluator data. The reproducibility archive includes the complete grader, fixtures and original records.

Historical disagreement-prediction benchmark

The earlier 13-model comparison predicts whether a panel will disagree; it does not ask models to complete the original job. Its precision, recall, F1 and weighted efficiency scores do not establish node suitability. All prior results remain available in full results.

Load outcomes_2026_10_13 for the 4,533 / 1,349 / 381 train / validation / test split. Labels are 0 = attested and 1 = disagreed, not factual correctness. Both validation and test have been evaluated; no fine-tuning was performed. The default outcomes configuration retains the original splits; keep their assignments separate.

Full source archive · Split metadata · Audit · Checksums

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