{"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":"sh0ck0r/Qwen3.5-0.8B-heretic","parameters":852985920,"revision":"fb35b60374fc15c99a3890e55a5a77e6d015cf4e","source":"https://huggingface.co/sh0ck0r/Qwen3.5-0.8B-heretic/tree/fb35b60374fc15c99a3890e55a5a77e6d015cf4e","updated":1792212205}],"model":"sh0ck0r/Qwen3.5-0.8B-heretic","page":"/models/sh0ck0r/Qwen3.5-0.8B-heretic","pending":[],"results":[{"benchmark":"ARC-Challenge / Qwen3.5 same-base cohort","coverage":"full","direction":"max","evaluation_seconds":39.85693414998241,"finished":1792212205,"gpu_slot_seconds":41.54360213293694,"manifest_url":"/artifacts/38365bef2439fd792f653a37/manifest.json","metric":"accuracy","n":1172,"protocol_version":"codesota-qwen35-arc-challenge-v1","revision":"fb35b60374fc15c99a3890e55a5a77e6d015cf4e","run_id":"38365bef2439fd792f653a37","run_url":"/runs/38365bef2439fd792f653a37","scope":"8877f1f73ee0a3c138937e07b7f6ca917f2c1d914d223e6a8ffa42b9ac8f0a47","score":0.32849829351535836,"task":"qwen35-arc-challenge-full"},{"benchmark":"ARC-Easy / Qwen3.5 same-base cohort","coverage":"full","direction":"max","evaluation_seconds":74.87046095100231,"finished":1792208915,"gpu_slot_seconds":76.55646423005965,"manifest_url":"/artifacts/338ec139014cbab8a49c1869/manifest.json","metric":"accuracy","n":2376,"protocol_version":"codesota-qwen35-arc-easy-v1","revision":"fb35b60374fc15c99a3890e55a5a77e6d015cf4e","run_id":"338ec139014cbab8a49c1869","run_url":"/runs/338ec139014cbab8a49c1869","scope":"84fe51c6790b19376ad4a234c742b80e0f73e3eb0d6a7257d48e5315d6a7c078","score":0.6519360269360269,"task":"qwen35-arc-easy-full"}],"revisions":["fb35b60374fc15c99a3890e55a5a77e6d015cf4e"],"source":"https://huggingface.co/sh0ck0r/Qwen3.5-0.8B-heretic"}}