{"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":"unsloth/Qwen3.5-0.8B-Base","parameters":852985920,"revision":"44bc982576bf30f0a9440303598859d92604ea5d","source":"https://huggingface.co/unsloth/Qwen3.5-0.8B-Base/tree/44bc982576bf30f0a9440303598859d92604ea5d","updated":1792212314}],"model":"unsloth/Qwen3.5-0.8B-Base","page":"/models/unsloth/Qwen3.5-0.8B-Base","pending":[],"results":[{"benchmark":"ARC-Challenge / Qwen3.5 same-base cohort","coverage":"full","direction":"max","evaluation_seconds":40.00580815796275,"finished":1792212314,"gpu_slot_seconds":41.77459662896581,"manifest_url":"/artifacts/55331fe6eafd22267a0df920/manifest.json","metric":"accuracy","n":1172,"protocol_version":"codesota-qwen35-arc-challenge-v1","revision":"44bc982576bf30f0a9440303598859d92604ea5d","run_id":"55331fe6eafd22267a0df920","run_url":"/runs/55331fe6eafd22267a0df920","scope":"8877f1f73ee0a3c138937e07b7f6ca917f2c1d914d223e6a8ffa42b9ac8f0a47","score":0.3848122866894198,"task":"qwen35-arc-challenge-full"},{"benchmark":"ARC-Easy / Qwen3.5 same-base cohort","coverage":"full","direction":"max","evaluation_seconds":74.56106622098014,"finished":1792208808,"gpu_slot_seconds":76.34397564595565,"manifest_url":"/artifacts/2d79cde7c16d073d6bb045a4/manifest.json","metric":"accuracy","n":2376,"protocol_version":"codesota-qwen35-arc-easy-v1","revision":"44bc982576bf30f0a9440303598859d92604ea5d","run_id":"2d79cde7c16d073d6bb045a4","run_url":"/runs/2d79cde7c16d073d6bb045a4","scope":"84fe51c6790b19376ad4a234c742b80e0f73e3eb0d6a7257d48e5315d6a7c078","score":0.70496632996633,"task":"qwen35-arc-easy-full"}],"revisions":["44bc982576bf30f0a9440303598859d92604ea5d"],"source":"https://huggingface.co/unsloth/Qwen3.5-0.8B-Base"}}