From 6c087cb5e2185879fc7fe0426336b95e5fb65275 Mon Sep 17 00:00:00 2001 From: autoresearch Date: Sun, 8 Mar 2026 16:27:48 +0000 Subject: [PATCH] revert interrupted softcap experiment (restore softcap=15) Co-Authored-By: Claude Opus 4.6 --- train.py | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/train.py b/train.py index c6c377b..668914a 100644 --- a/train.py +++ b/train.py @@ -278,7 +278,7 @@ class GPT(nn.Module): x = block(x, ve, cos_sin, self.window_sizes[i]) x = norm(x) - softcap = 17 + softcap = 15 logits = self.lm_head(x) logits = logits.float() logits = softcap * torch.tanh(logits / softcap)