{"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":"Qwen/Qwen3.5-0.8B","parameters":852985920,"revision":"2fc06364715b967f1860aea9cf38778875588b17","source":"https://huggingface.co/Qwen/Qwen3.5-0.8B/tree/2fc06364715b967f1860aea9cf38778875588b17","updated":1792212257}],"model":"Qwen/Qwen3.5-0.8B","page":"/models/Qwen/Qwen3.5-0.8B","pending":[],"results":[{"benchmark":"ARC-Challenge / Qwen3.5 same-base cohort","coverage":"full","direction":"max","evaluation_seconds":34.954991050995886,"finished":1792212257,"gpu_slot_seconds":36.617844913038425,"manifest_url":"/artifacts/4abfe34647b05095b1ea2fb3/manifest.json","metric":"accuracy","n":1172,"protocol_version":"codesota-qwen35-arc-challenge-v1","revision":"2fc06364715b967f1860aea9cf38778875588b17","run_id":"4abfe34647b05095b1ea2fb3","run_url":"/runs/4abfe34647b05095b1ea2fb3","scope":"8877f1f73ee0a3c138937e07b7f6ca917f2c1d914d223e6a8ffa42b9ac8f0a47","score":0.33276450511945393,"task":"qwen35-arc-challenge-full"},{"benchmark":"ARC-Easy / Qwen3.5 same-base cohort","coverage":"full","direction":"max","evaluation_seconds":76.11161479807924,"finished":1792209428,"gpu_slot_seconds":77.74721144407522,"manifest_url":"/artifacts/c73b968c1de14cc28b7e1717/manifest.json","metric":"accuracy","n":2376,"protocol_version":"codesota-qwen35-arc-easy-v1","revision":"2fc06364715b967f1860aea9cf38778875588b17","run_id":"c73b968c1de14cc28b7e1717","run_url":"/runs/c73b968c1de14cc28b7e1717","scope":"84fe51c6790b19376ad4a234c742b80e0f73e3eb0d6a7257d48e5315d6a7c078","score":0.6586700336700336,"task":"qwen35-arc-easy-full"}],"revisions":["2fc06364715b967f1860aea9cf38778875588b17"],"source":"https://huggingface.co/Qwen/Qwen3.5-0.8B"}}