From: Svjatoslav Agejenko <svjatoslav@svjatoslav.eu>
Date: Thu, 3 Jul 2025 20:05:52 +0000 (+0300)
Subject: New experimental inference settings.
X-Git-Url: http://www2.svjatoslav.eu/gitweb/?a=commitdiff_plain;h=62f2f8d1179055c21678647796186d5e4afac69f;p=alyverkko-cli.git

New experimental inference settings.
---

diff --git a/src/main/java/eu/svjatoslav/alyverkko_cli/AiTask.java b/src/main/java/eu/svjatoslav/alyverkko_cli/AiTask.java
index fc71786..c431b23 100644
--- a/src/main/java/eu/svjatoslav/alyverkko_cli/AiTask.java
+++ b/src/main/java/eu/svjatoslav/alyverkko_cli/AiTask.java
@@ -239,9 +239,8 @@ public class AiTask {
                 "--model " + mailQuery.model.filesystemPath,
                 "--threads " + configuration.getThreadCount(),
                 "--threads-batch " + configuration.getBatchThreadCount(),
-                "--top-k 1", "--top-p 0", // Activates true greedy decoding, now the recommended way to force determinism in recent llama.cpp builds
+                "--top-k 5", "--top-p 0.3",
                 "--min-p 0",
-                "--samplers \"top_k\"",
                 "--repeat-penalty 1.05",
                 "--dry-multiplier 0.4", // Low‐ish dry-multiplier adds a soft anti-repetition guard without wrecking logic
                 "--presence-penalty 0.1",
@@ -249,7 +248,7 @@ public class AiTask {
                 "--no-display-prompt",
                 "--no-warmup",
                 "--flash-attn",
-                "--temp 0",
+                "--temp 0.1",
                 "--ctx-size " + mailQuery.model.contextSizeTokens,
                 "--batch-size 512",
                 "--no-conversation",