straylight: cap Ornith 1.5 sampling and thinking budget

Default temp 0.6 / top-p 0.95 / top-k 20 and reasoning-budget 8192 so unrestricted <think> does not stall long agent turns.
2026-09-18 14:28:49 -07:00
parent 10dddaab7a
commit fdf50e0606
1 changed files with 6 additions and 0 deletions
@@ -73,8 +73,14 @@ let

    ; Ornith 1.5 35B-A3B MoE (official GGUF). ~3B active, multimodal; mmproj
    ; comes with --hf-repo. Q4_K_M is ~22 GB.
+    ; Sampling + a thinking cap: unrestricted <think> on this checkpoint stalls
+    ; long agent turns (Talos). Clients that send their own temperature still win.
    [ornith-1.5-35b-a3b]
    hf-repo = ornith-ai/Ornith-1.5-35B-A3B-GGUF:Q4_K_M
+    temp = 0.6
+    top-p = 0.95
+    top-k = 20
+    reasoning-budget = 8192
    dedup-cache-models = true

    ; Qwen 3.8 27B dense, refusal-ablated. Preferred uncensored coding agent.