{"api_version":"codesota.v1","data":{"archived_results":[],"evidence_status":"CodeSOTA measurements only; model-card claims are not benchmark results","measured_parameters":[{"basis":"Actual unique parameter count from pinned model loaded in CodeSOTA FP32 live-BPB evaluation","model":"SlayerLab/Slayer139-1.01b","parameters":139279760,"revision":"283ffb6d620be2d6ea1f2429d721962dc8e73633","source":"https://huggingface.co/SlayerLab/Slayer139-1.01b/tree/283ffb6d620be2d6ea1f2429d721962dc8e73633","updated":1792199477}],"model":"SlayerLab/Slayer139-1.01b","page":"/models/SlayerLab/Slayer139-1.01b","pending":[],"results":[{"benchmark":"ARC-Challenge","coverage":"subset","direction":"max","evaluation_seconds":2.01540831301827,"finished":1792205263,"gpu_slot_seconds":5.0236305829603225,"manifest_url":"/artifacts/400f0c30be43a9f62b75dd21/manifest.json","metric":"accuracy","n":32,"protocol_version":"codesota-sub1b-choice-v1","revision":"283ffb6d620be2d6ea1f2429d721962dc8e73633","run_id":"400f0c30be43a9f62b75dd21","run_url":"/runs/400f0c30be43a9f62b75dd21","scope":"e14efec58a56b426d1075a48d781531cc86194c0ae545ccae25ab8c742c4cdfa","score":0.28125,"task":"sub1b-arc-challenge"},{"benchmark":"ARC-Easy","coverage":"subset","direction":"max","evaluation_seconds":2.025514440028928,"finished":1792199493,"gpu_slot_seconds":4.873346148058772,"manifest_url":"/artifacts/81740f9404ca2161b03fab45/manifest.json","metric":"accuracy","n":32,"protocol_version":"codesota-sub1b-choice-v1","revision":"283ffb6d620be2d6ea1f2429d721962dc8e73633","run_id":"81740f9404ca2161b03fab45","run_url":"/runs/81740f9404ca2161b03fab45","scope":"44c19edbe693799c7f6c36c5dcae3a6094da38e830a1757f4ce21a6d173e9af6","score":0.375,"task":"sub1b-arc-easy"},{"benchmark":"ARC-Easy (full test split, raw accuracy)","coverage":"full","direction":"max","evaluation_seconds":71.80770976399072,"finished":1792206137,"gpu_slot_seconds":75.66824239201378,"manifest_url":"/artifacts/f54108c75de3695f938482e8/manifest.json","metric":"accuracy","n":2376,"protocol_version":"codesota-sub1b-arc-easy-standard-v2","revision":"283ffb6d620be2d6ea1f2429d721962dc8e73633","run_id":"f54108c75de3695f938482e8","run_url":"/runs/f54108c75de3695f938482e8","scope":"cf50d4b44de811db4aed984c74f66a1eab461af452494ae6c905f9c80d3ffc25","score":0.5471380471380471,"task":"sub1b-arc-easy-full"},{"benchmark":"BoolQ","coverage":"subset","direction":"max","evaluation_seconds":1.544806924997829,"finished":1792199580,"gpu_slot_seconds":6.976226528990082,"manifest_url":"/artifacts/ff71c6e3b5a06ff21384b00e/manifest.json","metric":"accuracy","n":32,"protocol_version":"codesota-sub1b-choice-v1","revision":"283ffb6d620be2d6ea1f2429d721962dc8e73633","run_id":"ff71c6e3b5a06ff21384b00e","run_url":"/runs/ff71c6e3b5a06ff21384b00e","scope":"bea6197d37b7692e8660c09f3be4bdc84c65bd6d240a6e23f418cc1d58d63ca5","score":0.75,"task":"sub1b-boolq"},{"benchmark":"HellaSwag","coverage":"subset","direction":"max","evaluation_seconds":2.026681552059017,"finished":1792199410,"gpu_slot_seconds":4.92196170205716,"manifest_url":"/artifacts/75ea5d8ae94de8a86a7a0777/manifest.json","metric":"accuracy","n":32,"protocol_version":"codesota-sub1b-choice-v1","revision":"283ffb6d620be2d6ea1f2429d721962dc8e73633","run_id":"75ea5d8ae94de8a86a7a0777","run_url":"/runs/75ea5d8ae94de8a86a7a0777","scope":"b4821eeee94ae676351fa0a7023207ccfb0c15fd7747debe3ffa69d04edf6456","score":0.375,"task":"sub1b-hellaswag"},{"benchmark":"OpenBookQA","coverage":"subset","direction":"max","evaluation_seconds":1.9651672700420022,"finished":1792205202,"gpu_slot_seconds":4.822887591901235,"manifest_url":"/artifacts/1a93c5498bda553a7673f19c/manifest.json","metric":"accuracy","n":32,"protocol_version":"codesota-sub1b-choice-v1","revision":"283ffb6d620be2d6ea1f2429d721962dc8e73633","run_id":"1a93c5498bda553a7673f19c","run_url":"/runs/1a93c5498bda553a7673f19c","scope":"14346d6c23a8cd6b6720f20dbe91c6ebdbfbb7721023028eb9705c7ca8a51fa7","score":0.25,"task":"sub1b-openbookqa"}],"revisions":["283ffb6d620be2d6ea1f2429d721962dc8e73633"],"source":"https://huggingface.co/SlayerLab/Slayer139-1.01b"}}