johnsonchromia's picture
Add Qwen results on 204 oracle protocol compliance cases
5eff9d4 verified
|
Raw History Blame Contribute Delete
4.22 kB

IMD oracle protocol compliance

204 derived cases from 102 real source requests, spanning five recipe kinds. Qwen3.8-27B completed both tracks: construction 0/102, repair 16/102. All 188 failures returned the recipe as a string rather than the required JSON object. This is a prompt-specific formatting result. Run report.

This tests submission construction and error repair. It does not test whether an oracle answer is true, whether the model obtained evidence, or whether it can build contracts. The original recipe and a supplied result are deliberately visible: this is a structured instruction-following task, not a hidden-answer reasoning test.

Track Cases Model task Pass condition
Construction 102 Package the supplied result and recipe into an IMD submission Pinned official checker passes; all required values and recipe are preserved
Repair 102 Fix a deliberately corrupted submission using the brief and checker error Same checks, without changing the supplied result

The five recipe kinds are call-compare (20 sources), log-sum (20), log-count (28), log-rank (16) and univ4-spot (18). Of 107 source requests with a recipe definition, four lacked explicit JSON and one used a filter unsupported by the pinned log-rank schema. All exclusions are recorded.

Repair corruptions are synthetic and labeled. They cover wrong request IDs, chains, windows or answer types; missing recipes; and unsupported fields. These are not claimed to be historical model mistakes. Every reference passes, every malformed control fails, and every changed-result control fails. Historical results are supplied inputs, not truth labels.

Construction and repair variants share source groups. Report pass rates separately and group related variants when estimating uncertainty or making splits. The tasks are public historical data and highly templated; 204 cases are not 204 independent job types. Preserve the 20-task chain-reading pilot as a separate execution track. Do not combine this suite into a general node-suitability score.

Files and scoring

  • tasks.jsonl: model inputs and provenance. Send only input plus the frozen system prompt.
  • labels.jsonl: expected artifact fields for the evaluator.
  • protocol.json: prompt, dataset hashes, scope and source counts.
  • preflight.json: positive and negative control results for every case.
  • exclusions.jsonl: excluded source requests and reasons.

The grader uses the pinned official check-answer.mjs in benchmarks/node-v1/vendor/, with its SHA-256 checked before use. It also checks exact JSON value preservation for v, requestId, chainId, window, answerType, answer and recipe. Valid optional notes or definitions are allowed. The task explicitly asks for the supplied JSON recipe, so alternative recipe formulations are outside this formatting test.

Run

python3 scripts/fetch_node_helpers.py
python3 scripts/run_protocol_benchmark.py --model qwen-27b --run-tag first --concurrency 2

The runner reads QWEN_API_KEY from the local .env and uses ModelScope with thinking disabled. It also supports sol and opus through the configured local gateway at medium effort. Each task gets one API attempt, temperature 0, at most 4,096 output tokens and a 120-second API timeout. All failures stay in the denominator. Output must be JSON without fences. Existing run directories cannot be overwritten.

Report construction and repair success separately, together with full task time, total recorded tokens, tokens per success, and missing usage. This is a model-serving comparison, not a native worker run. The frozen protocol.json records the pre-run snapshot (model_runs: 0); the completed run is recorded separately in the report. The prompt names the required fields but provides no formal output schema or worked example. No JSON-schema constrained decoding or interactive retries were used.

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

Python 3.9+ and Node 22+ are required. No JSON-schema constrained decoding or interactive correction retries were used. The frozen prompt is included in the metadata.