{"api_version":"codesota.v1","data":{"archived_results":[],"evidence_status":"CodeSOTA measurements only; model-card claims are not benchmark results","measured_parameters":[{"basis":"Unique parameters counted in the loaded evaluator model on White","model":"Mapika/decider-0.8b","parameters":752393024,"revision":"a0a01d6f8135298f400a8c856b355793012ae971","source":"https://huggingface.co/Mapika/decider-0.8b/tree/a0a01d6f8135298f400a8c856b355793012ae971","updated":1792212102}],"model":"Mapika/decider-0.8b","page":"/models/Mapika/decider-0.8b","pending":[],"results":[{"benchmark":"ARC-Challenge / Qwen3.5 same-base cohort","coverage":"full","direction":"max","evaluation_seconds":30.76483834197279,"finished":1792212102,"gpu_slot_seconds":32.45979036309291,"manifest_url":"/artifacts/0e0d5b30a00095f1910c3513/manifest.json","metric":"accuracy","n":1172,"protocol_version":"codesota-qwen35-arc-challenge-v1","revision":"a0a01d6f8135298f400a8c856b355793012ae971","run_id":"0e0d5b30a00095f1910c3513","run_url":"/runs/0e0d5b30a00095f1910c3513","scope":"8877f1f73ee0a3c138937e07b7f6ca917f2c1d914d223e6a8ffa42b9ac8f0a47","score":0.46757679180887374,"task":"qwen35-arc-challenge-full"},{"benchmark":"ARC-Easy / Qwen3.5 same-base cohort","coverage":"full","direction":"max","evaluation_seconds":87.03467097005341,"finished":1792208859,"gpu_slot_seconds":88.64412831701338,"manifest_url":"/artifacts/30d5530bf0ae38f549f7f381/manifest.json","metric":"accuracy","n":2376,"protocol_version":"codesota-qwen35-arc-easy-v1","revision":"a0a01d6f8135298f400a8c856b355793012ae971","run_id":"30d5530bf0ae38f549f7f381","run_url":"/runs/30d5530bf0ae38f549f7f381","scope":"84fe51c6790b19376ad4a234c742b80e0f73e3eb0d6a7257d48e5315d6a7c078","score":0.7575757575757576,"task":"qwen35-arc-easy-full"}],"revisions":["a0a01d6f8135298f400a8c856b355793012ae971"],"source":"https://huggingface.co/Mapika/decider-0.8b"}}