From 49d6daa11ecd4d9f237be53420c456a476b7c073 Mon Sep 17 00:00:00 2001
From: xaedes <xaedes@gmail.com>
Date: Sun, 7 May 2023 19:46:05 +0200
Subject: [PATCH] vastly improve training results

instead of logit targets 0 and 1 use -1 and +1.
---
 examples/baby-llama/baby-llama.cpp | 3 ++-
 1 file changed, 2 insertions(+), 1 deletion(-)

diff --git a/examples/baby-llama/baby-llama.cpp b/examples/baby-llama/baby-llama.cpp
index 0fbb01d5d..e5c548a05 100644
--- a/examples/baby-llama/baby-llama.cpp
+++ b/examples/baby-llama/baby-llama.cpp
@@ -1058,7 +1058,8 @@ void get_example_targets(int example_id, struct ggml_tensor * tokens_input, stru
     int n_tokens = tokens_input->ne[0];
     int n_vocab = targets->ne[0];
     float randomness = 0.0f;
-    ggml_set_zero(targets);
+    // ggml_set_zero(targets);
+    ggml_set_f32(targets, -1.0f);
     ggml_set_i32_1d(tokens_input, 0, 0);
     for (int i=1; i<n_tokens+1; ++i) {
         float x = example_id + i * 3.14159f * 2.0f * 1.0f / n_tokens;