{"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":"ahmedeltawel/Qwen3.5-0.8B-hsb-dsb","parameters":752393024,"revision":"7e781a3e970474d3fcb215ec0d3c5832e1131b9f","source":"https://huggingface.co/ahmedeltawel/Qwen3.5-0.8B-hsb-dsb/tree/7e781a3e970474d3fcb215ec0d3c5832e1131b9f","updated":1792212151}],"model":"ahmedeltawel/Qwen3.5-0.8B-hsb-dsb","page":"/models/ahmedeltawel/Qwen3.5-0.8B-hsb-dsb","pending":[],"results":[{"benchmark":"ARC-Challenge / Qwen3.5 same-base cohort","coverage":"full","direction":"max","evaluation_seconds":31.7163281890098,"finished":1792212151,"gpu_slot_seconds":33.369909026892856,"manifest_url":"/artifacts/153cea4d5c27ad0e88414281/manifest.json","metric":"accuracy","n":1172,"protocol_version":"codesota-qwen35-arc-challenge-v1","revision":"7e781a3e970474d3fcb215ec0d3c5832e1131b9f","run_id":"153cea4d5c27ad0e88414281","run_url":"/runs/153cea4d5c27ad0e88414281","scope":"8877f1f73ee0a3c138937e07b7f6ca917f2c1d914d223e6a8ffa42b9ac8f0a47","score":0.32593856655290104,"task":"qwen35-arc-challenge-full"},{"benchmark":"ARC-Easy / Qwen3.5 same-base cohort","coverage":"full","direction":"max","evaluation_seconds":83.22030307794921,"finished":1792209200,"gpu_slot_seconds":84.92086026701145,"manifest_url":"/artifacts/8400f70711b8e59ffcf9ecf7/manifest.json","metric":"accuracy","n":2376,"protocol_version":"codesota-qwen35-arc-easy-v1","revision":"7e781a3e970474d3fcb215ec0d3c5832e1131b9f","run_id":"8400f70711b8e59ffcf9ecf7","run_url":"/runs/8400f70711b8e59ffcf9ecf7","scope":"84fe51c6790b19376ad4a234c742b80e0f73e3eb0d6a7257d48e5315d6a7c078","score":0.6359427609427609,"task":"qwen35-arc-easy-full"}],"revisions":["7e781a3e970474d3fcb215ec0d3c5832e1131b9f"],"source":"https://huggingface.co/ahmedeltawel/Qwen3.5-0.8B-hsb-dsb"}}