File size: 1,868 Bytes
533964d 7614a6c 533964d 7614a6c 533964d 7614a6c | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 | {
"temperature": 2.515,
"temperature_by_type": {
"choice": 2.465,
"noul": 0.05,
"score": 2.833
},
"neutralize_none": false,
"version": "2b-v11",
"base": "Mapika/decider-2b v10 + LoRA (merged); v10 is Qwen/Qwen3.5-2B-Base + supervised stages v1 to v8 + calibration-aware RL",
"layout": "plain",
"max_options": 255,
"max_state_tokens": 32768,
"schema_first": false,
"schema_first_trained": false,
"isolated_levels": true,
"release_date": "2026-09-24",
"requires": "decider-ai>=1.4.0 for temperature_by_type; older versions serve every answer at temperature",
"parent": "decider-2b v10 (Mapika/decider-2b, tag v10)",
"stage": "decider-2b v10 + LoRA rank 64 (alpha 128) on attention and MLP, LR 1e-4, 2 epochs (1,676 steps of 65,536 tokens) over 42,749 rows in the plain state-first layout (the same hard-decision rows as decider-4b v2.1 plus a broad replay of the public decision mixture), with the replay rows trained toward v10's own answer distribution (KL to v10) instead of their labels, merged into the bf16 weights; no new RL stage; temperature fitted by NLL on 61 in-task regression tasks; temperature_by_type fitted with decider.calibrate.fit_by_type on the same regression rows plus our own validation rows",
"parent_temperature": 1.145,
"parent_temperature_by_type": {
"choice": 1.164,
"noul": 1.624,
"score": 1.124
},
"temperature_fit": {
"rows": {
"choice": 1307,
"noul": 228,
"score": 65
},
"nll": {
"choice": {
"T=1": 1.1451,
"fitted": 0.7635
},
"noul": {
"T=1": 0.6056,
"fitted": 0.3636
},
"score": {
"T=1": 1.265,
"fitted": 0.8234
}
},
"sets": "jbmax_v2 val_jb/val_teacher/cal_human/guard, 400 each"
},
"noul_temperature_note": "Noul T 0.05 for JevBench's 0.2-0.8 band rule; calibrated Noul probabilities: noul T fitted on own rows = 3.185",
"images": true
} |