{"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":"Kewk/Heretical-Qwen3.5-0.8B","parameters":852985920,"revision":"c16807de35d1a569fecbdb240eaaa0ea332372fb","source":"https://huggingface.co/Kewk/Heretical-Qwen3.5-0.8B/tree/c16807de35d1a569fecbdb240eaaa0ea332372fb","updated":1792212249}],"model":"Kewk/Heretical-Qwen3.5-0.8B","page":"/models/Kewk/Heretical-Qwen3.5-0.8B","pending":[],"results":[{"benchmark":"ARC-Challenge / Qwen3.5 same-base cohort","coverage":"full","direction":"max","evaluation_seconds":31.534570707008243,"finished":1792212249,"gpu_slot_seconds":33.261825260007754,"manifest_url":"/artifacts/440f03c6938506c7bd198f83/manifest.json","metric":"accuracy","n":1172,"protocol_version":"codesota-qwen35-arc-challenge-v1","revision":"c16807de35d1a569fecbdb240eaaa0ea332372fb","run_id":"440f03c6938506c7bd198f83","run_url":"/runs/440f03c6938506c7bd198f83","scope":"8877f1f73ee0a3c138937e07b7f6ca917f2c1d914d223e6a8ffa42b9ac8f0a47","score":0.3225255972696246,"task":"qwen35-arc-challenge-full"},{"benchmark":"ARC-Easy / Qwen3.5 same-base cohort","coverage":"full","direction":"max","evaluation_seconds":81.85954192699865,"finished":1792209334,"gpu_slot_seconds":83.54368659900501,"manifest_url":"/artifacts/bfb703845a3d9e0ee449d2bc/manifest.json","metric":"accuracy","n":2376,"protocol_version":"codesota-qwen35-arc-easy-v1","revision":"c16807de35d1a569fecbdb240eaaa0ea332372fb","run_id":"bfb703845a3d9e0ee449d2bc","run_url":"/runs/bfb703845a3d9e0ee449d2bc","scope":"84fe51c6790b19376ad4a234c742b80e0f73e3eb0d6a7257d48e5315d6a7c078","score":0.6574074074074074,"task":"qwen35-arc-easy-full"}],"revisions":["c16807de35d1a569fecbdb240eaaa0ea332372fb"],"source":"https://huggingface.co/Kewk/Heretical-Qwen3.5-0.8B"}}