From 00e64bd481aaafcac25ac3a89f166f7786fe8025 Mon Sep 17 00:00:00 2001 From: Anton Spilnyy Date: Fri, 21 Nov 2025 19:32:49 +0000 Subject: [PATCH] [aia-eval] LME-681 Configure eval for codex+gemini in AIA GitOrigin-RevId: 2e5beeb7968b2fab572c5a259d6623da87d7dbc4 --- .../codegeneration/java_swe_codex.json | 38 +++++++++++++++++++ .../codegeneration/java_swe_gemini.json | 38 +++++++++++++++++++ 2 files changed, 76 insertions(+) create mode 100644 plugins/evaluation-plugin/languages/evaluationconfig/codegeneration/java_swe_codex.json create mode 100644 plugins/evaluation-plugin/languages/evaluationconfig/codegeneration/java_swe_gemini.json diff --git a/plugins/evaluation-plugin/languages/evaluationconfig/codegeneration/java_swe_codex.json b/plugins/evaluation-plugin/languages/evaluationconfig/codegeneration/java_swe_codex.json new file mode 100644 index 000000000000..6abc0c979ec9 --- /dev/null +++ b/plugins/evaluation-plugin/languages/evaluationconfig/codegeneration/java_swe_codex.json @@ -0,0 +1,38 @@ +{ + "projectPath": "../ml-eval/jackson-core", + "repository": "https://github.com/fasterxml/jackson-core", + "revision": "f1dc3c512d211ae3e14fb59af231caebf037d510", + "outputDir": "bin/ml-eval-code-generation-swe-java-output", + "strategy": { + "instanceId": "fasterxml__jackson-core-1142", + "runner": "CODEX", + "invocationParams": { + "runFeatureTimeoutMin": 20 + } + }, + "fileDataset": { + "url": "remote:https://huggingface.co/datasets/JetBrains/eval_plugin/resolve/main/code-generation-swe/java/java_swe_bench.json" + }, + "interpret": { + "saveLogs": true, + "sessionProbability": 1.0, + "sessionSeed": null, + "sessionsLimit": 400, + "filesLimit": 100, + "order": "LINEAR", + "trainTestSplit": 70, + "registry": "", + "iterationCount": 1 + }, + "reports": { + "evaluationTitle": "Default", + "sessionsFilters": [], + "comparisonFilters": [] + }, + "language": "Java", + "actions": { + "evaluationRoots": [ + "." + ] + } +} diff --git a/plugins/evaluation-plugin/languages/evaluationconfig/codegeneration/java_swe_gemini.json b/plugins/evaluation-plugin/languages/evaluationconfig/codegeneration/java_swe_gemini.json new file mode 100644 index 000000000000..f2a17f2eba1d --- /dev/null +++ b/plugins/evaluation-plugin/languages/evaluationconfig/codegeneration/java_swe_gemini.json @@ -0,0 +1,38 @@ +{ + "projectPath": "../ml-eval/jackson-core", + "repository": "https://github.com/fasterxml/jackson-core", + "revision": "f1dc3c512d211ae3e14fb59af231caebf037d510", + "outputDir": "bin/ml-eval-code-generation-swe-java-output", + "strategy": { + "instanceId": "fasterxml__jackson-core-1142", + "runner": "GEMINI", + "invocationParams": { + "runFeatureTimeoutMin": 20 + } + }, + "fileDataset": { + "url": "remote:https://huggingface.co/datasets/JetBrains/eval_plugin/resolve/main/code-generation-swe/java/java_swe_bench.json" + }, + "interpret": { + "saveLogs": true, + "sessionProbability": 1.0, + "sessionSeed": null, + "sessionsLimit": 400, + "filesLimit": 100, + "order": "LINEAR", + "trainTestSplit": 70, + "registry": "", + "iterationCount": 1 + }, + "reports": { + "evaluationTitle": "Default", + "sessionsFilters": [], + "comparisonFilters": [] + }, + "language": "Java", + "actions": { + "evaluationRoots": [ + "." + ] + } +}