| { |
| "temperature": 1.145, |
| "temperature_by_type": { |
| "choice": 1.164, |
| "noul": 1.624, |
| "score": 1.124 |
| }, |
| "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" |
| } |
|
|
