From: Svjatoslav Agejenko <svjatoslav@svjatoslav.eu>
Date: Wed, 5 Nov 2025 23:44:38 +0000 (+0200)
Subject: Update hyperparameters for Tongyi DeepResearch
X-Git-Url: http://www2.svjatoslav.eu/gitweb/?a=commitdiff_plain;h=refs%2Fheads%2Fmaster;p=alyverkko-cli.git

Update hyperparameters for Tongyi DeepResearch
---

diff --git a/src/main/java/eu/svjatoslav/alyverkko_cli/AiTask.java b/src/main/java/eu/svjatoslav/alyverkko_cli/AiTask.java
index 9bb8bb7..07dff43 100644
--- a/src/main/java/eu/svjatoslav/alyverkko_cli/AiTask.java
+++ b/src/main/java/eu/svjatoslav/alyverkko_cli/AiTask.java
@@ -240,23 +240,23 @@ public class AiTask {
                 "--threads " + configuration.getThreadCount(),
                 "--threads-batch " + configuration.getBatchThreadCount(),
 
-                "--top-k 20",   // Restricts token selection to the K tokens with the highest probabilities.
+                //"--top-k 20",   // Restricts token selection to the K tokens with the highest probabilities.
 
                 "--top-p 0.95", // Restricts token selection to the smallest possible set
                 // of tokens whose cumulative probability exceeds the specified
                 // threshold P.
 
-                "--min-p 0.1",  // Filters the vocabulary to include only tokens whose
+                //"--min-p 0.1",  // Filters the vocabulary to include only tokens whose
                 // probability is at least a certain fraction (Min P) of the
                 // probability of the most likely token.
 
                 // "--chat-format qwen3",   // Ensure that model sees the <|im_start|>system … / <|im_start|>user … markup it was trained on
 
                 // Avoid getting stuck in a forever repetition loop
-                "--repeat-penalty 1.05", // Very little penalty, because computer code is often repetitive
+                "--repeat-penalty 1.1", // Very little penalty, because computer code is often repetitive
                 "--repeat-last-n 512",   // Last n tokens to consider for penalizing repetition
 
-                "--dry-multiplier 0.1", // Controls the strength of the penalty for a detected repetition sequence.
+                //"--dry-multiplier 0.1", // Controls the strength of the penalty for a detected repetition sequence.
 
                 "--presence-penalty 0", // In a code we want the model to reuse the same variable names,
                 // keywords, and syntax consistently. A presence penalty,
@@ -268,7 +268,7 @@ public class AiTask {
                 "--no-display-prompt",
                 "--no-warmup",
                 "--flash-attn on",
-                "--temp 0.6",
+                "--temp 0.85",
                 "--ctx-size " + mailQuery.model.contextSizeTokens,
                 "--batch-size 512",
                 "--single-turn", // run conversation for a single turn only, then exit when done